Videre Design STH-DCAM User manual

Small Vision System User Manual
Small Vision System
User’s Manual
Software version 2.3i
September 2002
? Kurt Konolige and David Beymer
SRI International
Small Vision System User Manual
1 Introduction ______________________________________________________ 5
1.1 The SRI Stereo Engine and the Small Vision System ________________________ 6
1.2 The Small Vision System_______________________________________________ 7
1.3 Hardware and Software Requirements ___________________________________ 8
Analog Framegrabbers ___________________________________________________ 8
Digital Framegrabbers ___________________________________________________ 8
1.4 The SVS Distribution _________________________________________________ 9
2 Getting started with smallv _______________________________________ 10
2.1 Inputting Live Stereo Video ___________________________________________ 12
Stereo heads __________________________________________________________
Analog Framegrabbers __________________________________________________
IEEE 1394 (FireWire) Framegrabber_______________________________________
Selecting Devices ______________________________________________________
Frame Size ___________________________________________________________
Image Sampling _______________________________________________________
Image Source _________________________________________________________
Streaming Mode _______________________________________________________
Adjusting Video Parameters______________________________________________
Subwindowing ______________________________________________________
Vergence ___________________________________________________________
Color ______________________________________________________________
2.2 Storing, Saving, and Loading Stereo Data________________________________ 21
Stereo Video Storage ___________________________________________________ 21
Loading and Storing Files _______________________________________________ 22
2.3 Display ____________________________________________________________ 23
2.4 Stereo Processing and Parameters ______________________________________ 24
Stereo Function________________________________________________________
3D Transformation _____________________________________________________
Calibration ___________________________________________________________
Disparity Search Range _________________________________________________
Adjusting the Horopter __________________________________________________
Pixel Information ______________________________________________________
Correlation Window Size ________________________________________________
Multiscale Disparity ____________________________________________________
2.5 Filtering ___________________________________________________________ 29
Confidence Filter ______________________________________________________ 29
Left/Right Filter _______________________________________________________ 29
2.6 Saving and Restoring Parameters ______________________________________ 30
3 Stereo Geometry__________________________________________________ 32
3.1 Disparity___________________________________________________________ 33
3.2 Horopter___________________________________________________________ 35
3.3 Range Resolution ____________________________________________________ 38
Small Vision System User Manual
3.4 Area Correlation Window_____________________________________________ 39
3.5 Multiscale Disparity__________________________________________________ 41
3.6 Filtering ___________________________________________________________ 42
3.7 Performance ________________________________________________________ 44
4 Calibration ______________________________________________________ 45
4.1 Calibration Procedure ________________________________________________ 46
Calibration procedure steps ______________________________________________ 46
Calibration Target _____________________________________________________ 48
Imager Characteristics __________________________________________________ 48
5 API Reference – C++ Language_____________________________________ 49
5.1 Threading and Multiple Stereo Devices __________________________________ 50
Threading Issues_______________________________________________________ 50
Multiple Devices_______________________________________________________ 50
5.2 C++ Classes ________________________________________________________ 51
5.3 Parameter Classes ___________________________________________________ 53
Class svsImageParams __________________________________________________ 53
Class svsRectParams ___________________________________________________ 53
Class svsDispParams ___________________________________________________ 53
5.4 Stereo Image Class __________________________________________________ 54
Constructor and Destructor ______________________________________________
Stereo Images and Parameters ____________________________________________
Rectification Information ________________________________________________
Disparity Image _______________________________________________________
3D Point Array ________________________________________________________
File I/O ______________________________________________________________
Copying Functions _____________________________________________________
5.5 Acquisition Classes __________________________________________________ 58
Constructor and Destructor ______________________________________________
Rectification __________________________________________________________
Controlling the Image Stream ____________________________________________
Error String __________________________________________________________
5.6 Video Acquisition____________________________________________________ 60
Video Object __________________________________________________________
Device Enumeration ____________________________________________________
Opening and Closing ___________________________________________________
Image Framing Parameters ______________________________________________
Image Quality Parameters _______________________________________________
Controlling the Video Stream ____________________________________________
5.7 File and Memory Acquisition __________________________________________ 63
File Image Object ______________________________________________________
Setting Images from Files________________________________________________
Stored Image Object ____________________________________________________
Setting Images from Memory_____________________________________________
5.8 Stereo Processing Classses ____________________________________________ 65
Stereo and 3D Processing ________________________________________________ 65
Small Vision System User Manual
Multiscale Stereo Processing _____________________________________________ 65
5.9 Window Drawing Classes _____________________________________________ 66
Class svsWindow ______________________________________________________ 66
Class svsDebugWin ____________________________________________________ 67
Small Vision System User Manual
The SRI Stereo Engine is an efficient realization of an area correlation algorithm for computing range
from stereo images. Figure 1 shows the results of running the algorithm on a typical scene. The image on
the top left is the left image of an original stereo pair, while the one on the top right is a disparity image
computed from the stereo pair. In the disparity images, brighter pixels show where the projection of an
object diverges between the images (has a high disparity). These are areas that are closer to the cameras.
Dark areas have lower disparity, and are further away. Finally, the bottom right shows a view of the 3D
reconstruction made from the disparity image.
Figure 1-1. An input image and the resultant stereo disparity image. Brighter areas are closer to
the camera.
Small Vision System User Manual
The SRI Stereo Engine and the Small Vision System
The Stereo Engine exists in several implementations, including embedded, low-power systems and
general purpose microcomputers. The embedded systems, or Small Vision Modules (SVMs), contain
DSPs or other standalone processors, and produce digital range information. They are meant for end
applications where size, cost, and power limitations are critical. SRI will develop embedded SVM
systems in partnership with companies who are interested in a particular application.
The Small Vision System (SVS) is an implementation of the Stereo Engine on general-purpose
microcomputers, especially PCs running Linux or Windows 95/98/ME/2000/NT. It consists of a set of
library functions implementing the stereo algorithms. Users may call these functions to compute stereo
results on any images that are available in the PC’s memory. Typically, standard cameras and video
capture devices are used to input stereo images. The Small Vision System is a development environment
for users who wish to explore the possibility of using stereo in an application.
This manual is useful as a source of general information about the Stereo Engine for any
implementation, but is also specifically aimed at the development environment of the SVS. It explains the
core characteristics of the Stereo Engine, serves as a reference for the stereo function API, and discusses
sample applications that use the API. There are also several tutorials that illustrate writing programs to
the SVS API, in the documentation folder. More technical information about stereo processing can be
found at, including several papers about the stereo algorithms and
With Version 2.2x of SVS, we introduce a re-written C++ interface to the SVS libraries. The new
C++ classes are much simpler to use than the previous C functions, and in particular relieve the user of
having to perform buffer management for images and disparity results. The C function interface is still
maintained for users who have an investment in application programs, under the 2.1x versions of SVS,
but we will only upgrade this version with bug fixes. Users are encouraged to migrate to the 2.2x API.
Small Vision System User Manual
The Small Vision System
The Small Vision System (SVS) is meant to be an accessible development environment for
experimenting with applications for stereo processing. It consists of a library of functions for performing
stereo correlation. Figure 1-2 shows the relationship between the SVS library and PC hardware.
Images come in via a pair of aligned video cameras, called a stereo head. A video capture board or
boards in the PC digitizes the video streams into main memory. The SVS functions are then invoked,
and given a stereo pair as an argument. These functions compute a disparity image, which the user can
display or process further.
Figure 1-2. The development environment of the Small Vision System.
The SVS environment of Figure 1-2 shows a typical setup for stereo processing of video images. The
user may supply his or her own cameras: the SVS has special processing for dealing with camera
distortion and calibrating the stereo image (Section 2.4.2). Special stereo heads are also available from
Videre Design ( The STH-V3 is an analog head with the ability to send a stereo
pair on just a single video signal, so only one video capture device is required (more information on video
capture is in Section 2.1). The MEGA-D (STH-MD1) is an all-digital device with megapixel imagers that
uses the 1394 bus (FireWire) for direct digital input. Finally, other sources of images may also be used, as
long as the images can be placed in PC memory. Some examples are images stored on disk, or images
obtained from other devices such as scanning electron microscopes.
Small Vision System User Manual
Hardware and Software Requirements
The SVS libraries exist for most Unix systems, as well as MS Windows 95/98/ME/2000/NT; that is,
on the most common computer platforms available. We have spent considerable effort in optimizing SVS
for PCs using the MMX instruction set (Pentium MMX and Pentium II, III, IV), and it will perform best
on these platforms, using either Linux or MS Windows. Performance information is in Section 3.7.
Analog Framegrabbers
Because of the ubiquity of PCs, we have added support for several PCI bus video capture devices on
PCs. The following are recommended PC hardware configurations for the SVS with analog cameras.
Operating System
Windows 95/98/ME
Windows NT, 2000
Video Capture Card
Imagenation PCX200
Matrox Meteor and Meteor RGB
Any Bt848-based card (e.g., Intel Smart Video Recorder III)
Imagenation PXC200
Matrox Meteor, Meteor RGB, Meteor II
Imagenation PXC200
Matrox Meteor, Meteor RGB, Meteor II
Table 1-1. Analog Framegrabbers and OS requirements for the Small Vision System.
Unfortunately, as of this time there are no good analog framegrabbers for laptops. The MRT Video
Port Pro is one of the best cards, but it is still slow for input to memory, and does not take advantage of the
32-bit CardBus specification. However, the MEGA-D digital device (see below) does have laptop input
Digital Framegrabbers
The MEGA-D (STH-MD1) and Dual DCAM (STH-DCAM) are all-digital devices that use the IEEE
1394 (FireWire) bus. Some desktops and laptops have 1394 ports integrated directly into their
motherboards. Otherwise, a standard 1394 PCI board or PCMCIA card can be used. The card must be
OHCI (Open Host Controller Interface) compliant, which almost all boards are.
Small Vision System User Manual
The SVS Distribution
The SVS distribution can reside in any directory; normally, it is placed in c:\svs (MS Windows
systems) or /usr/local/svs or a user’s directory (Unix systems). Here is the directory structure of
the SVS distribution.
svs.dll, lib
svscap.dll, lib
installation guide
release notes
PDF version of the User Manual
PDF version of the Calibration Addendum
executable and library files
full-featured GUI client demo
full-featured GUI client demo with calibration tool
simple stereo client example program
frame capture program, no stereo processing
framegrabber interface fns (Windows OS)
framegrabber interface fns (Linux)
SVS library (Windows OS)
SVS library for capture (Windows OS)
SVS library (Linux)
SVS library for capture (Linux)
Display library (Windows OS)
Display library (Linux)
stereo images
single-page printable calibration object
4-page printable calibration object
54mm single page calibration object (17x22)
108mm single page calibration object (34x44)
Sample stereo pairs and color files
Sample calibration files
sample client program sources
SVS library sources
main library header (C++)
framegrabber interface functions
Small Vision System User Manual
Getting started with smallv
The smallv program is a standalone application that exercises the SVS library. It is a GUI
interface to the stereo programs, and in addition can load and save stereo image sequences. The smallv
program is a useful tool for initial development of a stereo application, and can also be used to check out
and adjust a stereo camera setup.
The smallv program is in the bin/ directory. It requires shared libraries for the stereo
algorithms (svs), display (fltk), and calibration (various), all of which are in the bin/ directory.
Under MS Windows, these shared libraries (DLLs) must be in the same directory as the smallv
program, or in the system DLL directory. Under UNIX, the LD_LIBRARY_PATH variable must have the
path to the libraries.
Figure 2-1 shows the startup screen of the program. The black windows are for display of image and
stereo data. The display programs in SVS use the FLTK cross-platform window interface, and work best
in 24 bit video display mode. The version of the program is indicated in the text information area, and the
title bar.
smallv will accept stereo images from either a live video source, or a stored file. The easiest way
to get started with the program is to open a stored stereo sequence. From the File menu, choose Open,
and navigate to the data/ directory. The file face320-cal-X.bmp contains a stereo frame at
320x240 resolution. When you open it, it will show in the display windows. In the Function area,
pull down the list box and choose Stereo. Finally, press the Continuous button to compute the stereo
disparity and display it. You should see a green pattern representing stereo disparities in the right
window. Under the Horopter label, click the X offset button a few times to see the effect of
changing the stereo search area; a value of –4 or so should bring the close parts of the face into range.
Clicking the 3D Display button brings up an OpenGL window with a 3D view of the stereo data.
The rest of this section explains the operation of smallv. Since smallv exercises most of the
functionality of the SVS libraries, it should serve as a general introduction to the SVS functions. If you
Figure 2-1 Smallv program interface. The two black windows are for display of input images
and stereo results.
Small Vision System User Manual
are interested in using a particular framegrabber and set of cameras with smallv, please see Section 2.1.
The framegrabber interface is indicated in the message area on startup. In this case, it is the MEGA-D
digital stereo head.
Small Vision System User Manual
Inputting Live Stereo Video
The SVS libraries provide support for live video as stereo input. To input video, you must do the
following steps.
1. Decide on a stereo head and framegrabber.
2. Install the framegrabber, following instructions that come with the framegrabber or the stereo head
(STH-V3 or MEGA-D).
3. Copy the appropriate framegrabber DLL to bin\svsgrab.dll (MS Windows), or
bin/ (Unix); see Section 2.1.2).
4. Set the appropriate video format using the Video Format menu.
5. Set the video frame size.
6. Set the input mode to Video.
This section gives details necessary for performing these steps.
Stereo heads
Stereo requires two images from different viewpoints. The most common way to get these images is
to use two identical cameras separated by a horizontal baseline. It is important the cameras have lenses
with the same focal length, and that the pixel elements have the same size.
The baseline is typically from 3 to 8 inches wide, and the cameras are aligned parallel to each other,
signal 1
line of sight
signal 2
Figure 2-2 Stereo camera setup. Cameras are positioned with parallel
lines of sight. Their video signals are synchronized using cross-fed
although other configurations are possible. Figure 2-2 shows a typical stereo camera setup. Two cameras
are pointed in the same direction, and they are connected by a cable that genlocks the cameras, that is,
synchronizes them so that they capture images at the same time. Genlocking is important if there is any
motion in the scene. If the cameras are not genlocked, they can capture the image at slightly different
times, and any moving objects will be at a slightly different position in one camera relative to the other,
than if it they had taken the image at the same time. If the scene is static, then genlocking is not
necessary. Not every camera can be genlocked; check that the ones you have can be.
A word about monochrome vs. color cameras. If your application does not need color, it is preferable
to use monochrome cameras, because stereo relies only on the luminance component of the video signal.
Monochrome cameras have much better spatial resolution and dynamic range than color cameras of the
same quality, since they do not have to deal with three color channels. Having said this, the MEGA-D
megapixel cameras have such high resolution that using color imagers is generally not a problem, since
most applications can use 640x480 or 320x240 image sizes, and the color imagers produce excellent
quality by binning (averaging) a set of pixels.
The cameras produce two video outputs, which must be input to the PC running the SVS system.
There are three ways to get these video signals into the PC.
Small Vision System User Manual
1. Use two framegrabbers, and input one signal on each.
2. Use a single framegrabber capable of inputting 2 monochrome channels, e.g., the Matrox
Meteor RGB or Meteor II / Multichannel.
3. Use a single framegrabber, and a stereo head that interlaces two video signals onto a single
video stream. The STH-V3 from Videre Design ( is one such stereo
4. Use a digital stereo head, the MEGA-D (STH-MD1) from Videre Design. This stereo head
outputs a digital signal on the 1394 bus, and any OHCI (Open Host Controller Interface) card
can be used to input the video.
The SVS libraries can work with any size video frame up to 1288 by 1032 pixels. Standard NTSC
cameras capture frames up to 640 by 480, as a set of two fields, each 640 by 240. The camera first
captures a field in 1/60 of a second (the even field), then captures a second field 1/60 of a second later (the
odd field). The framegrabber can put these together to form a single image of 640 by 480 size. However,
the same problem with motion between non-genlocked cameras can occur on a single camera that
combines fields. There is a slight time delay between fields, leading to motion blur in the composed
frame. For this reason, the SVS libraries use fields rather than frames, so the maximum video size for
NTSC signals is 640 by 240.
PAL cameras are also support by the SVS libraries, as long as the framegrabber can input PAL video.
They involve similar considerations, but their maximum field size is 768 by 288.
Larger frame sizes with synchronized cameras are possible by using nonstandard progressive scan
analog cameras, or digital cameras.
Analog Framegrabbers
The SVS libraries include support for a number of popular analog signal framegrabbers; for digital
stereo heads, see the next subsection. The table below lists them according to their operating system.
MS Windows
MS Windows
NT 4.0
Matrox Meteor, Meteor RGB, Meteor PPB
Any Bt848-based card, e.g.
Intel Smart Video Recorder III
Imagenation PXC2000
Matrox Meteor, Meteor RGB, Meteor PPB
Matrox Meteor II
Imagenation PXC200
Matrox Meteor, Meteor RGB, Meteor PPB
Meteor II
Imagenation PXC200
MSW Installation
Table 2-1 Framegrabbers supported by SVS.
Under MS Windows, a particular framegrabber is accessed from the SVS libraries by copying the
corresponding DLL and LIB files. Copy the appropriate C++ library (.dll) for your framegrabber to
svsgrab.dll, and the corresponding reference file (.lib) to svsgrab.lib. For example, if you
have installed the Matrox Meteor II board, then copy the following files in the bin\ directory:
svsmet2.dll -> svsgrab.dll
svsmet2.lib -> svsgrab.lib
Small Vision System User Manual
There are convenient BAT files to set up a particular framegrabber, in the bin\ directory. Simply
double-click on the batch file to perform the required copying. For example, to set up the Imagenation
PXC200 framegrabber, go to the bin\ directory in Windows Explorer, and double-click on the file
The framegrabber interface must be set up before starting smallv; it cannot be changed while the
program is running.
Under Linux, the interface to either the Matrox Meteor cards or a Bt848 card (PXC200, Intel Smart
Video Recorder, etc.) is with the shared library
Copy this library to
bin/ to use it. You must load the proper low-level driver for the card, namely, the
BTTV drivers. These drivers are included with almost every distribution of Linux; we recommend the
current Red Hat distribution.
The framegrabber interface used by smallv under MS Windows or Linux is indicated in the
message window at startup.
There are some limitations in framegrabber drivers that should be noted. First, there are currently no
good analog framegrabbers for portables. For the digital interface, any standard IEEE 1394 OHCI card
can be used; this is one of the advantages of the digital interface.
IEEE 1394 (FireWire) Framegrabber
The SVS has an interface to digital stereo heads from Videre Design via the IEEE 1394 serial bus.
Any OHCI-compliant IEEE 1394 PCI or PCMCIA card can be used, under MS Windows 98/2000 or
Linux. Please check the stereo head manual for instructions on installing the 1394 card and drivers.
The relevant interface libraries are given in Table 2-2 below. To set up a particular interface in
Linux, copy the library file to bin/ Under MS Windows, execute the setup file in the
Stereo Head
MS Windows
MSW Installation File
Table 2-2 Interface libraries for Videre Design digital stereo heads.
bin\ directory by double-clicking on it in Windows Explorer.
Selecting Devices
The Device menu button lets you tell the SVS library what kind of video input you are using. For
Videre Design digital heads (MEGA-D and Dual DCAM), you can select among multiple devices attached
to any IEEE 1394 card on your computer. For analog stereo heads, you can choose how to input images
from the analog framegrabber boards. Currently, there is no way to select multiple analog stereo heads
connected to the same computer.
Digital Stereo Devices (MEGA-D and Dual DCAM)
A digital stereo head attached to the IEEE 1394 bus is recognized by smallv, and the Device menu
button will drop down a list of these devices when selected. Devices have a number, which starts at 1 for
the first device encountered. These numbers can change as devices are added or dropped from the bus.
Devices also have an id, which is a numeric string that is unique to the device. The Device list shows
both the current device number, and its unique id. The currently selected device is indicated by a checked
box; you can change the current device by selecting any available device. This choice becomes active the
next time Video input is selected in the input choice box.
Small Vision System User Manual
Left camera
camera field
Figure 2-3 Line-interlaced stereo cameras (STH-V3 stereo head). Alternate
lines from each camera are interlaced into a single video stream.
Only one type of device, the MEGA-D or Dual DCAMs, will be seen by the smallv program. The
choice depends on which interface library has been loaded (see Section 2.1.3). It is possible to mix these
devices on the same IEEE 1394 bus, but a given application will see only one type of device or the other.
Analog Stereo Devices (STH-V3 and user cameras)
Analog camera input goes through an analog framegrabber , where it is converted into digital form
and sent to the host computer memory. The smallv interface will input video from the cameras through
either one or two framegrabbers, depending on the type of setup. There are three choices:
1. Line interlace [default]. This is the mode for the STH-V3 line interlace stereo head. Any
framegrabber can be used in this mode.
2. Dual framegrabbers. This mode uses two framegrabbers, with one framegrabber per camera.
Check Table 2-1 for supported framegrabbers.
3. RG components. This mode uses the Matrox Meteor RGB or Meteor II, and inputs one
camera video stream on the R channel, and one on the G channel.
It is important for stereo processing to have the left camera image appear as the left image in the
smallv program. Once the video input is displayed, you can check this by pointing the cameras along
your line of sight. The right camera appears on your right side, and the right image on the smallv
display should show this image. You can cover one camera with your hand, and observe which displayed
image goes dark. With dual framegrabbers or RG input, the solution to having the wrong camera inputs
is to simply switch the inputs, or to use the swap button in smallv, which interchanges the images in
Under line interlace mode (Figure 2-3), the first horizontal line of a video field is from the left
camera, the second from the right, the third from the left, and so on, making a single video stream. The
SVS software de-interlaces the video stream, reconstructing the left and right images in memory, at half
the original vertical resolution. Because of the variation in how framegrabbers determine which is the
first line of a field, the SVS software will sometimes switch the left and right fields during de-interlacing.
The swap button switches the left and right fields during deinterlacing.
The smallv application interfaces only to one analog stereo head, using the first one or two
framegrabbers that it finds. It is possible to specify other framegrabbers from a use application, using the
Frame Size
The SVS libraries as delivered can work with frame sizes up to 1288 by 1032. In fact, the SVS
algorithms can work with arbitrarily sized frames.
Small Vision System User Manual
A subset of frame sizes are supported for video input in the smallv application; the following table
summarizes them. Most framegrabbers support hardware interpolation and scaling, so that bus traffic is
minimized by working with smaller frames. The exception is the Meteor RGB, which passes a full field
to memory, where it is decimated by the SVS software.
Video Format
Line interlace
Dual framegrabber
and RG component
1394 (digital) interface
Frame Sizes
1280x960, 640x480, all
others above
Table 2-3 Frame sizes available for video input
in smallv.
Video frame size is selected with the Size drop list in the Source area. Video size can be changed
only when frames are not being acquired. Once acquisition starts, the frame size is fixed.
Image Sampling
The sampling for analog framegrabbers is implicit in the frame size. For example, if the camera
image size is 320x240, and the requested frame size is 160x120, then the full image is scaled down by the
framegrabber, usually using interpolation to produce a smooth image.
With the MEGA-D digital interface, the user has full control over the sampling method, and the
Sample and Size controls combine to produce the final result. For example, if the sampling mode is x1
(no subsampling), then and image size of 320x240 produces a subwindow within the full image.
(Subwindowing is not available from analog framegrabbers supported by SVS.) The placement of the
subwindow can be changed in real time under program control, using the dialog from the Subwindow…
There are several sampling modes. Decimation samples the image by removing pixels, e.g., “x2
dec” means that every second pixel in a line is removed, and every other line is removed. Binning
samples the image by averaging over a block of four pixels, to produce the same result. Binning produces
smoother images with less noise, but it is slower than decimation, which is done by the stereo hardware.
Combination sampling modes are available, e.g., “x4 bin+dec” samples the image down to ¼ size in
horizontal and vertical directions, by decimating by 2 and then binning by 2.
The Dual DCAM stereo device supports a single sampling mode, x2 binning at 320x240 output
resolution. The binning mode will reduce video noise for this device.
Image Source
The source for stereo images can be either a memory buffer or a live video stream. The Source
drop list lets you choose between these, or to stop any input. Buffer input is discussed in Section 2.1.9.
Streaming Mode
Images from video cameras or the buffer can be processed in three acquisition modes. Only one
acquisition mode is active at a given time.
? ? Continuous mode. In this mode, stereo pairs are continuously input, processed, and
displayed. The maximum frame rate is 30 Hz for live analog image data, and up to 80 Hz for
the MEGA-D digital system. See Section 3.7 for performance information. The rate is
indicated next to the text information area.
? ? Single frame mode. In this mode, a single stereo pair is input, processed, and displayed each
time the Single button is pressed.
Small Vision System User Manual
? ? Freeze mode. In this mode, a single stereo pair is input, then the same frame is continuously
processed and displayed. This mode is useful in checking the effect of different stereo
parameters on the same image.
Adjusting Video Parameters
Most framegrabbers support some kind of video image adjustments, such as contrast or brightness.
The video parameter dialog is invoked using the Video… menu item (Figure 2-4).
Most analog cameras have automatic adjustment of exposure and gain, which change according to
lighting conditions. The user can set brightness and contrast, which are framegrabber parameters that
change the processing of the analog signal.
The MEGA-D digital stereo head has manually controlled exposure and gain. Exposure is the time
that any given pixel is exposed to light before being read out. Gain is a amplification of the signal that
comes out of the pixel. In general, it is best to increase the exposure first, and if necessary, to increase
gain once exposure reaches a maximum. The reason for this is that gain will increase the video noise,
while exposure increases the pixel’s response to light. In some cases, though, short exposure times are
desirable for minimizing motion blur, and it may be more convenient to increase gain while exposure is
not at a maximum.
The values of exposure, gain, brightness, and contrast are all represented as a percent.
The colorized version of the MEGA-D digital camera can input color images, and the color balance
can be adjusted manually using the red/blue differential gain. More information about color processing is
in Section 2.1.12.
Figure 2-4 Video Parameter dialog box.
Small Vision System User Manual
2.1.10 Subwindowing
The MEGA-D digital stereo head can send to the host computer just a portion, or subwindow, of the
stereo image. For example, if the MEGA-D is in x2 sampling mode (full-size image is 640x320), and the
image size is chosen to be 320x240, then smallv will input only a 320x240 subwindow of the full
image. Figure 2-5 shows two of these subwindows, and the original full-size image.
The placement of subwindows is controlled by the vertical (Y) and horizontal (X) offset controls in
the Subwindow dialog window; the dialog is initiated from the Subwindow… menu item in the main
window. These parameters can be changed in real time, enabling electronic panning of the live image.
Figure 2-5 Two 320x240 subwindows (bottom) of a 640x480 image (top).
Small Vision System User Manual
2.1.11 Vergence
When in subwindow mode, the two cameras in a stereo rig generally will have the same X and Y
offsets, so that they keep the parallel line-of-sight characteristic of the stereo rig. However, for viewing
close objects, it is advantageous to toe-in, or verge, the two stereo cameras. In this way, the images of the
near object will both contain the object in the center.
Human eyes verge mechanically when viewing close objects. Mechanical vergence for stereo cameras
is difficult, however, since it involves complicated motor control, and more importantly, disturbs the
calibration that is critical for stereo analysis. Instead, with the subwindow capability of the MEGA-D, it
is possible to verge the stereo images electronically, by choosing appropriate horizontal offsets for each
Figure 2-6 shows the effects of using electronic vergence. The top stereo pair, of a close object, puts
the object into the center of the left frame. In the right frame, the object has a large disparity and is
visible in the left side of the frame.
The bottom stereo pair is created by adding vergence to the subwindow process, offsetting the right
subwindow horizontally by 120 pixels, relative to the left subwindow. Both frames now have the near
object centered.
Vergence of the subwindows is set using the vergence control in the Subimage box of the
Subwindow dialog. It is a real time control, just like the X and Y subwindow offsets.
Figure 2-6 Parallel image subwindows (top) and verged image subwindows (bottom), showing a
close object.
Small Vision System User Manual
2.1.12 Color
As of Version 2.1, SVS supports color input and display. Besides the two monochrome left/right
stereo channels, there is a third color channel that corresponds to the left image, with images in RGB 32bit format, and optionally a fourth color channel for the right image. The color channels do not
participate in stereo processing, but can be useful in applications that combine color and stereo
information, for example, object tracking. Usually only the left color channel is needed, since the left
image is the reference image for stereo disparities and 3D information. Both color channels are available
for user applications, if desired. When the right color channel is requested, the left color channel is
always also provided.
Color information from the MEGA-D digital head (STH-MD1-C) is input as raw colorized pixels,
and converted by the interface library into two monochrome and one or two RGB color channels. The
main color channel corresponds to the left image, which is the reference image for stereo. The color
image can be de-warped, just like the monochrome image, to take into account lens distortion (Section 4).
Optionally, a second color channel is available for the right image.
The stereo DCAM device (STH-DCAM) performs color processing on-camera, and sends the results
down to the application. De-warping proceeds as for the MEGA-D.
Color information from the camera is input only if the Color button is pressed on the main window
(Figure 2-1)., under the appropriate window.
To get the color images in applications, use the
SetColor() command.
Because the typical color camera uses a colorizing filter on top of its pixels, the color information is
sampled at a lower resolution than a similar non-colorized camera samples monochrome information. In
general, a color camera has about ¼ the spatial resolution of a similar monochrome camera. To
compensate for the reduced resolution, use binning (Section 2.1.6) to increase the fidelity of the image.
For example, if you need a 320x240 frame size, use 640x480 and binning x2.
The relative amounts of the three colors, red/green/blue, affects the appearance of the color image.
Many color CCD imagers have attached processors that automatically balance the offsets among these
colors, to produce an image that is overall neutral (called white balance). The MEGA-D provides manual
color balance by allowing variable gain on the red and blue pixels, relative to the green pixels. Manual
balance is useful in many machine vision applications, because automatic white balance continuously
changes the relative amount of color in the image. The STH-DCAM allows for either automatic or
manual control of color (see Figure 2-4
The manual gain on red and blue pixels is adjusted using the Video Parameters window
(Section 2.1.9). For a particular lighting source, try adjusting the gains until a white area in the scene
looks white, without any color bias.
Small Vision System User Manual
Storing, Saving, and Loading Stereo Data
smallv provides a basic facility for loading and saving stereo data streams. The file load and store
functions are part of the SVS library, and their source code is included. smallv exercises these
functions, and provides a memory buffer for storing live stereo video.
There are two basic types of storage/playback available. The first, video storage, is meant to save a
video sequence of stereo images (including color information). Images from live video streaming are
captured to an internal buffer of 200 frames. These frames can be replayed and stored.
A second type of storage is still image storage. In this mode, a single frame is stored to a file. The
video storage buffer is not involved in this process, and an arbitrary number of such stills can be saved.
Stereo Video Storage
smallv has an internal buffer capable of holding 200 stereo pairs (frames). Depending on the size
of the images and the amount of memory on the machine, a video sequence of frames can be captured to
physical memory without slowing down video capture.
The buffer can be filled from a previously-saved file set, or from live video input. The buffer can also
be written out to a file set, and used as the source for stereo processing in smallv.
The video buffer is controlled from the Video Buffer window, accessed via the menu bar (Figure
When the input source is the buffer, the acquisition mode controls (Continuous, Single,
Freeze) control the processing of the buffer frames (Section 2.1.8). The frame control can also be used
to go to an individual frame when in Single acquisition mode.
The Record button controls the input of live video into the buffer. Clear clears the buffer and
resets it to frame 0. Activating the Record button starts the input of live video frames into the buffer.
The source must be set to Video; either Continuous or Single mode may be used. Frames are
stored sequentially until the buffer is full. Pressing Record again will also turn off acquisition before the
buffer is full.
As an example, to capture a short video sequence and replay it, perform the following steps.
1. Start acquiring live video in continuous mode.
2. Clear the buffer (Clear button).
3. Start buffer storage (Record button).
4. After a short period, stop buffer storage (Record button).
5. Change from Video to Buffer source.
At this point, the short segment that is in the buffer can be replayed as a short continuous loop. The
Figure 2-7 Video buffer controls.
Small Vision System User Manual
buffer, or individual images, can be saved to a file. Under the File menu, use the commands Store
Video Buffer and Load Video Buffer. A video sequence is stored as a set of BMP files (next
subsection), in a sequence starting at 001. You cannot save stereo disparity or 3D data directly from the
video buffer. However, the current frame disparity and 3D data can be saved using the still image storage
facility, described in the next subsection.
Loading and Storing Files
The SVS libraries work with different file types for image storage.
? ? BMP format. Each BMP file contains a single 8-bit grayscale image, or an RGB 24-bit color
image. The color coding for the 8-bit BMP file is 256 shades of gray, with 0 being black and
255 white. By convention, a stereo pair is saved as two files with the linked names XXXL.BMP (left image) and XXX-R.BMP (right image). The corresponding left color image is
saved as XXX-C.BMP, and the right color image as XXX-Q.BMP. Finally, the image
parameters are stored as a text file XXX.INI.
? ? Text files for disparity images. Disparity images can be saved as a text file, with one line of
text for each line of the image (e.g., a 320x240 image will have 240 lines). Each line
contains an image row of disparities, as integers. The special values –1 and –2 indicate that
the disparities were filtered out, by the texture measure (-1) or the left/right check (-2)..
? ? Text files for 3D points. 3D point arrays, generated from a disparity image (Section 2.4.2),
can be saved as a text file. Each line of the file represents one point of the array. The array
has the same format as the image from which it was produced, e.g., if the input image is
320x240, then the file has 320x240 lines, in row-primary order. Each line has 3D X,Y,Z
coordinates first, as floating-point numbers, then three integers for the R,G,B values of the
pixel at that point. If the disparity at a pixel is filtered out, then the Z value is negative,
indicating a filtered value.
Still images and still image sequences are loaded using the File menu. To load stereo frames, use
the Load Images (BMP) menu item to bring up a file choice dialog. Choosing either BMP file of a
pair automatically loads the other. In addition, if a color file or parameter file (.ini) is present, it is also
A file image sequence is a set of files with a base name XXXNNN. For example, CAL001-L.BMP is
the left image of a stereo pair in a sequence. The sequence can start with any number, but the number
must be 3 digits, and must increment sequentially for each stereo pair in the sequence. Choosing Single
from the smallv interface will load the next file in the sequence. Freeze reloads the same file
continuously, which is useful for changing stereo parameters in smallv and seeing their effects.
Continuous mode is not available when loading a sequence of still images.
Note that loading a still image file using the Load Images menu item does not load the images into
the video buffer. The only way to load images into the video buffer is with the Load Video Buffer
menu item.
To save the current stereo image to a file, use the Store Current (BMP) menu item. Color
information, if present, is saved as a 24-bit BMP file. A sequence of still images can be created by using
the correct format for the base name of the stored files, as described above.
If stereo processing is turned on and a disparity image has been produced, then it can be saved as an
8-bit BMP file (Store Disparity Image (BMP) ), or as a text file (Store Disparity Image
(Text)). Since the number of bits in the disparity pixels is generally greater than 8, the BMP file only
contains the high-order bit information. The text file contains the full value for the disparity.
3D information (X,Y,Z values) can be saved to a file, if stereo processing is selected. Use the Store
3D Point Array menu item.
Small Vision System User Manual
smallv displays two images in its display area. The left display is always the left input image.
Input images are displayed in grayscale, unless color information is present: in this case, the left image
will be shown in color.
The right display can be either the right input image, or the results of stereo processing. Processing
results are always displayed in “greenscale”, using shades of green.
Either display can be turned off by unchecking the box underneath the display area. Turning off the
display will let smallv run faster.
Images larger than 320x240 are automatically scaled down by factors of 2n to fit into a 320x240 area.
Smaller image sizes are displayed in the original size.
To display properly for human viewing, most video images are formatted to have a nonlinear
relationship between the intensity of light at a pixel and the value of the video signal. The nonlinear
function compensates for loss of definition in low light areas. Typically the function is x? , where ? is 0.45,
and the signal is called “gamma corrected.” Digital cameras, such as the MEGA-D, do not necessarily
have gamma correction. This is not a problem for stereo processing, but does cause the display to look
very dark in low-light areas. You can add gamma correction to the displayed image by choosing an
appropriate gamma value in the slider under the right display window (Figure 2-1).
Small Vision System User Manual
Stereo Processing and Parameters
In smallv, stereo processing takes place in conjunction with the input of stereo images. The basic
cycle is:
get stereo pair -> process pair -> display pair
The input is either from live video or the buffer (Sections 2.1 and 2.1.9). In freeze mode, the same pair is
processed continuously, so adjustments can be made in stereo parameters.
Stereo Function
Stereo processing is turned on by choosing Stereo from the Function drop list. The stereo
disparity image will appear in the right display. Stereo disparities are encoded by green: brighter green is
a higher disparity, and therefore closer to the cameras (see Section 2.4.4 for a technical description of
Disparities represent the distance between the horizontal appearance of an object in the stereo images.
The stereo process interpolates this distance to 1/16 pixel, e.g., a disparity value of 45 represents a
displacement of 2 13/16 pixels. The maximum displacement currently supported is 80 pixels, so disparity
values range from 0 (no disparity) to 1280. Disparity values are returned as 16-bit (short) integers. The
values 0xFFFF and 0xFFFE are reserved for filtering results (Section 2.5)
If smallv is running on an MMX processor (Pentium or AMD) then stereo processing is much
faster, taking advantage of the parallel data operations. The processor is queried and the MMX box is
checked if the instructions are available. You can turn the MMX processing on and off by toggling the
box. But, if your system does not have MMX instructions, you will not be able to turn it on.
3D Transformation
A pixel in the disparity image represents range to an object. This range, together with the position of
the pixel in the image, determines the 3D position of the object relative to the stereo rig. SVS contains a
function to convert disparity values to 3D points. These points can then be displayed in a 3D viewer.
To take the current disparity image and display it in 3D, press the 3D Display button. An
OpenGL window will show the 3D points constructed from the disparity image, and you can change the
viewpoint of the window to see the 3D structure (Figure 2-8).
The coordinate system for the 3D image is taken from the optic center of the left camera of the stereo
rig. Z is along the optic axis, with positive Z in front of the camera. X is along the camera scan lines,
positive values to the right when looking along the Z axis. Y is vertical, perpendicular to the scan lines,
with positive values down.
The X and Y position of the viewpoint, as well as rotation around the Z axis, can be changed with the
sliders on the left side of the window. The scale of the image can be changed as well. Finally, the
viewpoint can be rotated around a point in the image, to allow good assessment of the 3D quality of the
stereo processing. The rotation point is selected automatically by finding the point closest to the left
camera, near the optic ray of that camera. To rotate the image around this point, put the mouse in the 3D
window, and drag the pointer while holding the left button down.
Small Vision System User Manual
Figure 2-8 3D display window. The red ray is the optic ray from the left camera.
Small Vision System User Manual
For good stereo processing, the two images must be aligned correctly with respect to each other. The
process of aligning images is called calibration. Generally speaking, there are two parts to calibration:
internal calibration, dealing with the properties of the individual cameras and especially lens distortion;
and external calibration, the spatial relationship of the cameras to each other. Both internal and external
calibration are performed by an automatic calibration procedure described in Section 4. The procedure
needs to be performed when lenses are changed, or the cameras are moved with respect to each other.
From the internal and external parameters, the calibration procedure computes an image warp for
rectifying the left and right images. In stereo rectification, the images are effectively rotated about their
centers of projection to establish the ideal stereo setup: two cameras with parallel optical axes and
horizontal epipolar lines (see Fig. 2-2). Having the epipolar lines horizontal is crucial for correspondence
finding in stereo, as stereo looks for matches along horizontal scanlines.
Figure 2-9 shows a pair of images of the calibration target taken with the MEGA-D stereo head and a
4.8 mm wide-angle lens. In the original images on the top, there is lens distortion, especially at the edges
of the image: notice the curve in the target. Also, the images are not aligned vertically.
The bottom pair is the result of calibrating the stereo head and then rectifying the two original
images. Now the images are aligned vertically, and all scene lines are straight in the images.
Figure 2-10 shows sample disparity images for uncalibrated and calibrated cameras. Without
calibration, it is impossible for the stereo algorithms to find good matches.
Disparity Search Range
Even with stereo rectification, it may not be possible to match every object in the scene, because the
horopter is not large enough. In this case, the horopter can be enlarged by changing the number of
disparities searched by the stereo process. This search range can vary from 8 to 80 pixels. Larger search
Figure 2-9 Original stereo pair (top) and rectified pair (bottom).
Small Vision System User Manual
ranges enlarge the horopter, but not in a linear fashion, i.e., a search range of 32 does not give twice the
horopter range of 16; see Section 4 for technical details.
Changing the disparity search size affects the time it takes to process stereo. A search space of 32
pixels will take about twice as long as a search space of 16 pixels. It will actually take a little less,
because there is some fixed overhead in processing the images. Obviously, the smallest search range
necessary for the application is the best choice.
Disparities are interpolated to 1/16 pixel, so a search range of 16 means that there are 256 integral
disparity values, ranging from 0 (no disparity) to 255 (maximum disparity of 15 15/16 pixels).
The search range is selected using the Disparities value in the Parameters area. When the
range is switched, the disparity image will lighten or darken to reflect the changed values of disparities.
Adjusting the Horopter
The stereo rectification procedure sets up the horopter, or depth of field of stereo, so that objects are
matched from infinity to some distance in front of the camera. Objects closer than this near point will not
be matched, and will produce random disparity readings. The near point distance is a function of the
search size, the stereo baseline, and the focal length of the camera lenses. One can adjust the horopter by
adjusting a horizontal X offset, moving the depth range closer to the camera. The depth range desired in
the end application would drive the setting of this parameter. For example, if the image does not contain
any objects farther than a certain distance, the X offset can be adjusted so that the far point of the horopter
is at that distance. Changing the X offset causes the disparity display to get uniformly lighter or darker,
as the horopter is shifted and the disparity of an object changes. Adjusting the horopter to cover a specific
range of depths is discussed in Section 4.
Pixel Information
SVS will show pixel information when the left button is clicked in either SVS display window. The
information is displayed in the text window in the format:
x232 y120 [131] [11] Xaaa Ybbb Zccc
The image coordinates of the mouse are given by the x,y values. The values in square brackets are the
pixel values of the left and right images. If the right image is displaying stereo disparities, then the right
value is the disparity value. Finally, the X,Y,Z values are the real-world coordinates of the image point,
in mm. Note that X,Y,Z values are calculated only if stereo is being computed, and to be accurate, a
good calibration file must be input (Section 4).
Correlation Window Size
The size of the correlation window used for matching affects the results of the stereo processing. A
Figure 2-10 Uncalibrated (left) and calibrated (right) disparity images.
Small Vision System User Manual
larger window will produce smoother disparity images, but will tend to “smear” objects, and will miss
smaller objects. A smaller window will give more spatial detail, but will tend to be noisy. Typical sizes
for the window are 9x9 or 11x11. The window size is selected using the Sum window drop list. In the
MMX implementation, not all window sizes are supported. More technical information on the correlation
window can be found in Section 3.4.
Multiscale Disparity
Multiscale processing can increase the amount of information available in the disparity image, at a
nominal cost in processing time. In multiscale processing, the disparity calculation is carried out at the
original resolution, and also on images reduced by 1/2. The extra disparity information is used to fill in
dropouts in the original disparity calculation (Figure 3-8 in Section 2.4.8).
Multiscale processing is turned on in smallv by enabling the MultiScale button.
Small Vision System User Manual
Stereo processing will generally contain incorrect matches. There are two major sources for these
errors: lack of sufficient image texture for a good match, and ambiguity in matching when the correlation
window straddles a depth boundary in the image. The SVS stereo processing has two filters to identify
these mismatches: a confidence measure for textureless areas, and a left/right check for depth boundaries.
Areas that are filtered appear black in the displayed disparity image. To distinguish them from valid
disparity values, they have the special values 0xFFFF (confidence rejection) and 0xFFFE (left/right
Confidence Filter
The confidence filter eliminates stereo matches that have a low probability of success because of lack
of image texture. There is a threshold, the confidence threshold, that acts as a cutoff. Weak textures give
a confidence measure below the threshold, and are eliminated by the filter.
The confidence threshold is adjusted using the Conf spin control in the Parameters area. A
good value can be found by pointing the stereo cameras at a textureless surface such as a blank wall, and
starting the stereo process. There will be a lot of noise in the disparity display if the confidence threshold
is set to 0. Adjust the threshold until the noise just disappears, and is replaced by a black area.
The computational cost of the confidence filter is negligible, and it is usually active in a stereo
Left/Right Filter
Each stereo camera has a slightly different view of the scene, and at the boundaries of an object there
will be an area that can be viewed by one camera but not the other. Such occluded areas cause problems
for stereo matches. Fortunately, they can be detected by a consistency check in which matching is done
first by using the left image as a fixed base, and then repeating the match using the right image as the
base. Disparity values for the same point that are not the same fail the left/right check. Typically, this
will occur near the boundaries of objects.
The left/right check is controlled by three radio buttons in the Parameter area. It can be turned on
or off. A third option is to perform the check, but instead of discarding disparity values that are
inconsistent, use the one that is smaller (further away). This option can fill in the areas around object
borders in a reasonable way. It is not currently available under MMX processing.
The left/right check adds about 20% to the computational cost of the stereo process, but is usually
worth the effort.
Small Vision System User Manual
Saving and Restoring Parameters
All of the parameters that control the operation of the SVS Stereo Engine can be saved to a file for
later use. Parameter files can be loaded and saved using the File menu: Load Param File and
Store Param File.
The file data/megad-75.ini contains a sample file for a 7.5 mm lens on the MEGA-D stereo
rig. It serves as an example of the settings available through parameter files. In practice, these settings
are usually computed using the calibration program, and then saved to a file for later use. But, it is also
possible to change the settings directly in the file.
# SVS Engine v 2.2 Stereo Camera Parameter File
max_linelen 1280
max_lines 960
max_decimation 4
max_binning 2
gamma 0.700000
color 0
ix 0
iy 0
vergence 0
rectified 0
width 320
height 240
linelen 320
lines 240
decimation 2
binning 2
subwindow 1
have_rect 1
autogain 0
manualgain 1
autowhite 0
manualwhite 1
gain 0
exposure 100
contrast 0
brightness 50
saturation 20
red 0
blue 0
convx 9
convy 9
corrxsize 11
corrysize 11
thresh 20
lr 1
ndisp 24
dpp 16
offx 0
offy 0
# image frame parameters
# max size of imager
allowable decimation at imager
allowable binning in driver
gamma correction for display
0 for monochrome, 1 for color
subwindow offset
# vergence of right subwindow
# subwindow size
# window size
# current decimation and binning
1 for subwindow capability
1 if we have rectification parameters
1 if autogain available
1 if manual gain available
1 if auto white balance available
1 if manual white balance available
current gain value [0,100], neg for auto
current exposure [0,100], neg for auto
current contrast [0,100]
current brightness [0,100]
current saturation [0,100]
current red gain [-40,40], neg for auto
current blue gain [ -40,40]
# stereo processing parameters
# prefilter kernel size
# correlation window size
confidence threshold value
left/right filter on (1) or off (0)
number of disparities to search
subpixel interpolation
horopter offset
vertical image offset, not used
Small Vision System User Manual
[left camera]
pwidth 1280
pheight 960
dpx 0.007500
dpy 0.007500
sx 1.000000
Cx 582.260123
Cy 506.081223
f 7.798704
kappa1 0.002983
kappa2 -0.000040
# translation between left and right cameras
# rotation between left and right cameras
# number of pixels in the camera
# pixel spacing, mm
# aspect ratio
# camera center, pixels
# focal length, mm
# radial distortion parameters
# projection matrix: from left cam era 3D coords
to left rectified coordinates
1.041674e+003 6.177793e+000 5.666963e+002 0.000000e+000
-6.957139e+000 1.042596e+003 5.022628e+002 0.000000e+000
-6.576823e-003 -3.900478e-005 1.000000e+000 0.000000e+000
# rectification matrix for left camera
1.001883e+000 -5.935817e-003 1.956143e+001
6.693463e-003 1.000873e+000 1.397722e+000
6.313708e-006 1.844243e-015 1.000000e+000
[right camera]
pwidth 1280
pheight 960
dpx 0.007500
dpy 0.007500
sx 1.000000
Cx 548.992956
Cy 495.924832
f 7.834438
kappa1 0.002722
kappa2 -0.000021
# number of pixels in the camera
# pixel spacing, mm
# aspect ratio
# camera center, pixels
# focal length, mm
# radial distortion parameters
# projection matrix: from right camera 3D coords
to right rectified coordinates
1.041674e+003 6.177795e+000 5.666964e+002 -9.352453e+004
-6.957140e+000 1.042596e+003 5.022628e+002 4.191134e -004
-6.576824e-003 -3.900531e-005 1.000000e+000 -9.225858e-005
# rectification matrix for right camera
1.006349e+000 -9.720995e-003 -1.942293e+001
8.006440e-003 9.997475e-001 -1.474720e+000
9.887915e-006 -7.757015e-006 1.000000e+000
Small Vision System User Manual
Stereo Geometry
Stereo algorithms compute range information to objects by using triangulation. Two images at
different viewpoints see the object at different positions: the image difference is called disparity. This
section discusses the basic equations that govern the relationship between disparity and range.
Small Vision System User Manual
The figure below displays stereo geometry. Two images of the same object are taken from different
viewpoints. The distance between the viewpoints is called the baseline (b). The focal length of the lenses
is f. The horizontal distance from the image center to the object image is dl for the left image, and dr for
the right image.
Figure 3-1. Definition of disparity: offset of the image location of
an object.
Normally, we set up the stereo cameras so that their image planes are embedded within the same
plane. Under this condition, the difference between dl and dr is called the disparity, and is directly related
to the distance r of the object normal to the image plane. The relationship is:
(1) r = bf / d , where d = dl - dr .
Using Equation 1, we can plot range as a function of disparity for the STH-V1 stereo head. At their
smallest baseline, the cameras are about 8 cm apart. The pixels are 14 um wide, and the standard lenses
have a focal length of 6.3 mm. For this example, we get the plot in Figure 3-2. The minimum range in
this plot is 1/2 meter; at this point, the disparity is over 70 pixels; the maximum range is about 35 meters.
Because of the inverse relationship, most of the change in disparity takes place in the first several meters.
The range calculation of Equation (1) assumes that the cameras are perfectly aligned, with parallel
image planes. In practice this is often not the case, and the disparity returned by the Stereo Engine will be
offset from the ideal disparity by some amount X0. The offset is explained in the section below on the
horopter, and in the section on calibration.
Small Vision System User Manual
Figure 3-2. Inverse relationship between disparity and range. This plot
is for a focal length of 6.3 mm, a baseline of 80 mm, and a pixel
width of 14 mm.
Small Vision System User Manual
Stereo algorithms typically search only a window of disparities, e.g., 16 or 32 disparities. In this case,
the range of objects that they can successfully determine is restricted to some interval. The horopter is the
3D volume that is covered by the search range of the stereo algorithm. The horopter depends on the
camera parameters and stereo baseline, the disparity search range, and the X offset. Figure 3-3 shows a
typical horopter. The stereo algorithm searches a 16-pixel range of disparities to find a match. An object
Plane of
Plane of
16 disparities
Figure 3-3 Horopter planes for a 16-pixel disparity search.
that has a valid match must lie in the region between the two planes shown in the figure. The nearer
plane has the highest disparity (15), and the farthest plane has the lowest disparity (0).
The placement of the horopter can be varied by changing the X offset between the two images, which
essentially changes the search window for a stereo match. Figure 3-5 shows the raw disparities for a
typical stereo head. The cameras are slightly verged, so a zero disparity plane (where an object appears at
the same place in both images) occurs at some finite distance in front of the cameras. If the stereo
algorithm is searching 5 disparities, then without any X offset, it will search as shown in the top red
arrow, that is, from disparity 0 to disparity 4. By offsetting one image in the X direction by n pixels, the
horopter can be changed to go from –n to 5-n raw disparities. This search range is indicated by the lower
red arrow.
Generally, it is a good idea to set the X offset to compensate for camera vergence or divergence, that
is, to set it so that the furthest horopter plane is at infinity. The reason that this is a good idea is because
it’s usually possible to control how close objects get to the camera, but not how far away. The offset that
puts the far horopter plane at infinity is called X0. With this offset, a disparity of 0 indicates an infinitely
far object.
The horopter can be determined from Equation (1). For example, if the disparity search window is 031, the horopter (using the graph above) will be from approximately 1 meter to infinity. The search
window can be moved to an offset by shifting the stereo images along the baseline. The same 32 pixel
window could be moved to cover 10-41 pixel disparities, with a corresponding horopter of 0.8 meters to
2.2 meters.
Small Vision System User Manual
Disparity: 5 4 3 2 1
Figure 3-5. Planes of constant disparity for verged stereo cameras. A search range
of 5 pixels can cover different horopters, depending on how the search is offset
between the cameras.
The location and size of the horopter depends on the application. If an object falls outside the
horopter, then its true disparity will not be found, and instead it will get some random distribution of
disparities. Figure 3-4 shows what happens when the object's range falls outside the horopter. In the left
image, the disparity search window is correctly positioned so that objects from 1 meter to infinity are in
view. In the right image, the window has been moved back so that objects have higher disparities.
However, close objects are now outside of the horopter, and their disparity image has been "broken up"
into a random pattern. This is typical of the disparity images produced by objects outside the horopter.
For a given application, the horopter must be large enough to encompass the ranges of objects in the
application. In most cases, this will mean positioning the upper end of the horopter at infinity, and
making the search window large enough to see the closest objects.
The horopter is influenced not only by the search window and offset, but also by the camera
parameters and the baseline. The horopter can be made larger by some combination of the following:
? ? Decreasing the baseline.
? ? Decreasing the focal length (wider angle lenses).
? ? Increasing pixel width.
? ? Increasing the disparity search window size.
As the cameras are moved together, their viewpoints come closer, and image differences like disparity
are lessened. Decreasing the focal length changes the image geometry so that perceived sizes are smaller,
and has a similar effect. It also makes the field of view larger, which can be beneficial in many
applications. However, very small focal length lenses often have significant distortion that must be
corrected (see the section on calibration). Another way to change the image geometry is to make the pixels
wider. This can be done by scaling the image, e.g., from 320x240 to 160x120, which doubles the pixel
size. Note that it is only necessary to change the pixel width. Most framegrabbers have hardware scaling
Figure 3-4. Disparity image for all regions withing the horopter (left) and
some regions outside the horopter (right).
Small Vision System User Manual
to arbitrary resolutions.
These first three options change the camera geometry, and thus have a corresponding effect on the
range resolution, which decreases (see below). The only way to increase the horopter size and maintain
range resolution is to increase the disparity search window size, which leads to more computation.
Multiresolution methods, which use several sizes of an image, each with its own horopter, are one way to
minimize computation (see, for example, the paper by Iocchi and Konolige at
Small Vision System User Manual
Range Resolution
Often it's important to know the minimal change in range that stereo can differentiate, that is, the
range resolution of the method. Give the discussion of stereo geometry above, it's easy to see that that
range resolution is a function of the range itself. At closer ranges, the resolution is much better than
farther ranges.
Range resolution is governed by the following equation.
(2) ? r = (r2/bf) ? d
The range resolution, ? r, is the smallest change in range that is discernable by the stereo geometry,
given a change in disparity of ? d. The range resolution goes up (gets worse) as the square of the range.
The baseline and focal length both have an inverse influence on the resolution, so that larger baselines
and focal lengths (telephoto) make the range resolution better. Finally, the pixel size has a direct
influence, so that smaller pixel sizes give better resolution. Typically, stereo algorithms can report
disparities with subpixel precision, which also increases range resolution.
The figure below plots range resolution as a function of range for the STH-MD1 (MEGA-D) stereo
head, which has a baseline of 9 cm. The Stereo Engine interpolates disparities to 1/16 pixel, so ? d is 1/16
* 7.5 um = 0.08533 um. The range resolution is shown for a sampling of different lens focal lengths. At
any object distance, the range resolution is a linear function of the lens focal length.
Equation 2 shows the range resolution of a perfect stereo system. In practice, video noise, matching
errors, and the spreading effect of the correlation window all contribute to degrading this resolution.
Range resolution is not the same as range accuracy, which is a measure of how well the range
computed by stereo compares with the actual range. Range accuracy is sensitive to errors in camera
calibration, including lens distortion and camera alignment errors.
Figure 3-6. Range resolution as a function of range. This plot assumes a baseline of 90
mm, and a pixel size of 7.5 um, with subpixel resolution of 1/16 pixel.
Small Vision System User Manual
Area Correlation Window
Stereo analysis is the process of measuring range to an object based on a comparison of the object
projection on two or more images. The fundamental problem in stereo analysis is finding corresponding
elements between the images. Once the match is made, the range to the object can be computed using the
image geometry.
Area correlation compares small patches, or windows, among images using correlation. The window
size is a compromise, since small windows are more likely to be similar in images with different
viewpoints, but larger windows increase the signal-to-noise ratio. Figure 3-7 shows a sequence of
disparity images using window sizes from 7x7 to 13x13. The texture filter was turned off to see the
effects on less-textured areas, but the left/right check was left turned on.
There are several interesting trends that appear in this side-by-side comparison. First, the effect of
better signal-to-noise ratios, especially for less-textured areas, is clearly seen as noise disparities are
eliminated in the larger window sizes. But there is a tradeoff in disparity image spatial resolution. Large
windows tend to “smear” foreground objects, so that the image of a close object appears larger in the
disparity image than in the original input image. The size of the subject’s head grows appreciably at the
end of the sequence. Also, in the 7x7 the nose can be seen protruding slightly; at 13x13, it has been
smeared out to cover most of the face.
One of the hardest problems with any stereo algorithm is to match very small objects in the image. If
an object does not subsume enough pixels to cover an appreciable portion of the area correlation window,
it will be invisible to stereo processing. If you want to match small objects , you have to use imagers with
good enough spatial resolution to put lots of pixels on the object.
Small Vision System User Manual
Figure 3-7 Effects of the area correlation window size. At top is the original left
intensity image. The greenscale images show windows of 7x7, 9x9, 11x11,
and 13x13 windows (clockwise from upper left).
Small Vision System User Manual
Multiscale Disparity
Multiscale processing can increase the amount of information available in the disparity image, at a
nominal cost in processing time. In multiscale processing, the disparity calculation is carried out at the
original resolution, and also on images reduced by 1/2. The extra disparity information is used to fill in
dropouts in the original disparity calculation (Figure 3-8).
Figure 3-8 Effects of multiscale disparity calculation. Upper figure shows disparity dropouts in a
typical scene, where there is not enough texture for correlation to be reliable. Adding disparity
information from a ½ resolution image (lower part of figure) shows additional coverage in the
disparity image.
Small Vision System User Manual
Like most vision algorithms, the results of stereo processing can contain errors. In the case of stereo,
these errors result from noisy video signals, and from the difficulty of matching untextured or regularly
textured image areas. Figure 3-9 shows a typical disparity image produced by the SRI algorithm. Higher
disparities (closer objects) are indicated by brighter green (or white, if this paper is printed without color).
There are 64 possible levels of disparity; in the figure, the closest disparities are around 40, while the
furthest are about 5. Note the significant errors in the upper left and right portion of the image, where
uniform areas make it hard to estimate the disparity.
In Figure 3-9(c), the interest operator is applied as a postfilter. Areas with insufficient texture are
rejected as low confidence: they appear black in the picture. Although the interest operator requires a
threshold, it’s straightforward to set it based on noise present in the video input. Showing a blank gray
area to the imagers produces an interest level related only to the video noise; the threshold is set slightly
above that. Or, more simply, you can use the temporal variance of poorly textured matches to adjust the
texture threshold. Observing the disparity image during realtime display, there will usually be areas that
flicker rapidly. Adjust the threshold upward until these regions disappear. If there are no such regions,
adjust the threshold downward until just before they appear.
(a) Input grayscale image, one of a stereo pair
(c) Texture filter applied
(b) Disparity image from area correlation
(d) Left/right and texture filter applied
Figure 3-9 Post-filters applied to a disparity image. (c) is a texture filter that eliminates
textureless areas. (d) is a consistency check between left and right stereo matches.
There are still errors in portions of the image with disparity discontinuities, such as the side of the
subject’s head. These errors are caused by overlapping the correlation window on areas with very
different disparities. Application of a left/right check can eliminate these errors, as in Figure 3-9(d). The
Small Vision System User Manual
left/right check can be implemented efficiently by storing enough information when doing the original
disparity correlation.
In practice, the combination of an interest operator and left/right check has proven to be the most
effective at eliminating bad matches.
Small Vision System User Manual
Using standard PC hardware, running either MS Windows 95/98/ME/2000/XP/NT or Linux, the SVS
can compute stereo range in real time. Table 3-1 gives some typical timings for a 500 MHz Pentium III
processor. Because the Stereo Engine has a very small memory footprint, the timings scale almost
linearly with increasing processor speed. These timings include the complete stereo algorithm detailed
above: dewarping of input images, disparity computation and interpolation, and post-filtering using a
texture filter and left/right filter.
Frame size
Number of
Frame Rate
180 Hz
100 Hz
45 Hz
24 Hz
6 Hz
Table 3-1 Processing rates on a Pentium III
500 MHz machine.
Small Vision System User Manual
NOTE: There is a Calibration Addendum manual that details the exact steps necessary to
perform calibration, and includes troubleshooting information. Please consult that manual for more
detailed information about the calibration procedure.
Most stereo camera setups differ from an ideal setup in which the cameras are perfect pinhole imagers
and are aligned precisely parallel. The divergence from ideal causes problems in the quality of the stereo
match since epipolar lines are not horizontal. In addition, if the camera calibration is unknown, one does
not know how to interpret the stereo disparities in terms of range to an object. Camera calibration
addresses these issues by creating a mathematical model of the camera.
SVS incorporates a simple automatic procedure for calibration, using a planar object that can be
printed on a standard printer. The calibration is preformed by fitting a model to a number of images
taken of a planar calibration object. The user presents the object to the stereo rig in five different
(arbitrary) poses. The calibration procedure finds model features in the images, and then calculates a
best-fit calibration for the rig. The procedure works for many different combinations of imagers,
baselines, and lenses, including wide-angle lenses with severe distortion.
When is it necessary to perform calibration? In general, whenever an action changes the camera
intrinsics (lens focal length and center axis) or extrinsics (the cameras move with respect to each other).
Here are some actions that would necessitate re-calibration:
? ? Changing lenses
? ? Screwing the lenses in or out of their mount
? ? Zooming, if the lenses are zoom lenses
? ? Changing the baseline of the cameras
? ? Any movement or rotation of one camera independent of the other, e.g., severe vibration or shock
can change the cameras’relative position
A rigid mount that keeps the cameras stable with respect to each other is a necessity for a stereo rig. For
example, the MEGA-D uses an extruded aluminum frame to stabilize the cameras. There are some
actions that do not require re-calibration:
? ? Changing the lens focus with a focusing ring on the lens
? ? Changing the lens aperture
The next section reviews the calibration procedure, detailing the steps required to generate a
calibration file.
Small Vision System User Manual
Calibration Procedure
An automatic calibration procedure using five image pairs of a planar calibration target is included as
part of the smallv program. Given the image pairs of the calibration object, the system automatically
locates corner features in the target, fits a model of the target to the images, and finally produces an
estimate for the left and right camera intrinsics, the stereo head extrinsics, and the rectification matrices
P0 , P1 , H 0 , and H1 . These values are then used by smallv, and can also be saved as a parameter file
for later re-use. More information about the calibration procedure can be found in the Calibration
Addendum to the User’s Manual.
Calibration procedure steps
Create the calibration object. Print out a copy of the file data/check.pdf (Figure 4-1) and
paste it to a surface that is as flat as possible. We use a wooden cutting board as a backing
surface at SRI.
Start the smallv application and start capturing video. It is recommended that you set the video
resolution to at least 320x120 in order to get enough detail of the calibration object. A calibration
computed when capturing video at a higher resolution can be used for future video captured at
any resolution with the same cameras.
Bring up the calibration window by pressing the Calibrate… menu button. Fig. 4-4 shows the
calibration dialog window (the figure shows the dialog after an image has been captured and
Determine the appropriate characteristics of the camera imagers and enter them into the four
boxes in the middle of the dialog. If you have one of the Videre Design stereo heads, check the
appropriate box and the parameters are loaded automatically.
Acquire five stereo pairs of the calibration object at different rotations and translations. Try to
avoid views that differ by a simple translation, as they are less informative than views with
variation in rotation. As shown in Fig 4-4, there is a tab control that shows only one pair at a
time; choose a tab to select another pair. To capture the current video feed into a stereo pair box,
simply press the capture button. You can also save and load images to and from disk using the
load and save buttons.
Figure 4-1 Checkerboard calibration object.
Small Vision System User Manual
Figure 4-4 Smallv calibration dialog window.
Detect the corner features in all views by pressing the features button in the lower bar of
buttons. This cycles through all the images, displaying the corners in green as they are detected.
If the feature finder fails on an image, please re-capture the image and redetect the features.
When redetecting features for a single stereo pair, use the features button in the stereo pair
7. Compute the calibration parameters (intrinsics, extrinsics) and rectification matrices by pressing
the calibrate button in the lower button bar. This operates in three phases:
a) Calibrates individual views using a planar model of the calibration object. The projection of
these model features is shown in red.
b) Calibrates all the views jointly using nonlinear optimization over all the intrinsic and
extrinsic parameters. This phase usually takes a few minutes, and when finished, the
projected model features are shown in yellow.
c) Computes the rectification matrices from the joint calibration in (b).
8. When the calibration is finished, the parameter listing at the bottom is updated, and you have a
couple additional options in the lower button bar: save writes the parameter file to disk and OK
exits the calibration dialog and transfers the new parameters to the main smallv window.
Small Vision System User Manual
Calibration Target
The standard calibration target, check.pdf, can be printed out on a single piece of 8.5 x 11 inch
paper. In some cases, this image is too small, e.g., when using wide baseline cameras. There is another
target, check2.pdf, that has squares twice as large as the standard one (54 mm instead of 27 mm).
This target must be assembled from four printed pieces of 8.5 x 11 inch paper. It is recommended for
large-baseline systems, where the target must be placed at a distance from the cameras in order to be seen
in both of them.
Imager Characteristics
The calibration process must be related to the geometry of the camera imagers. There are three
important parameters:
1. Pixel size in mm (width and height)
2. The width of the imager in pixels
3. The width of the image output by the framegrabber
The stereo rigs from Videre Design (MEGA-D, STH-V2, and STH-V3) have preset values; just use the
appropriate button in the Calibrate dialog. Otherwise, select the Custom value, and choose values
according to the instructions below.
The pixel size can be found from the specifications of the imager. If you cannot find these, just use
defaults of 0.010 mm. The calibration will determine the pixel aspect ratio (width / height). The
calibrated lens focal length will not be correct because the pixel scale will be off, but that will not change
the validity of the calibration.
The imager width is the number of sels (sensor elements) in each line of the imager. Again, this can
be found in the imager specifications. If these are not available, just use the image width as it comes from
the framegrabber, e.g., for NTSC video it is 640.
The framegrabber width is the width of the image output by the framegrabber, in pixels. Typically it
will be an analog NTSC signal, which is 640 pixels. For digital imagers, such as the MEGA-D, the image
size in sels and the framegrabber width are the same.
Small Vision System User Manual
API Reference – C++ Language
With SVS 2.2x, the standard programming interface to the SVS libraries is in C++. To add stereo
processing to your own programs, you call functions in the Stereo Engine library. These functions are
available in svs.dll (Windows 95/98/2000/NT) or (Unix systems). The header file is
src/svsclass.h. The current version of the library is 2.2d.
Source code samples for the C++ API are in the directory samples/. A simple example of the use
of these functions is in the sample program samples/stframe.cpp.
Small Vision System User Manual
Threading and Multiple Stereo Devices
Threading Issues
The SVS core library functions (svs.dll, are thread-safe: they can be used in any
thread in a process. Of course, the user is responsible for not overlapping calls in different threads, e.g.,
starting up two competing disparity calculations using the same object in different threads.
The MEGA-D and Dual-DCAM acquisition libraries are also thread-safe, in general. However, there
are some known quirks under MS Windows. The most obvious of these is the Open() call for the
MEGA-D. This call must be made in the main thread. Subsequent accesses using GetImage() can be
made in any thread.
Graphic window output is handled by the FLTK cross-platform windowing system. This system is
not, in general, thread safe. Calls to the FLTK functions must all be made from the same thread; multiple
threads are allowed in the application program, as long as all FLTK calls come from the same thread.
There is a nascent thread locking mechanism in FLTK, but it is not yet incorporated into SVS.
Multiple Devices
Multiple stereo devices (MEGA-D and Dual-DCAM) can be accessed simultaneously from a single
process, or from multiple processes. Only one process may access a given device, using an Open() call.
Once this call is made, subsequent calls to Open() from other processes will fail until the device is
Currently, there are several restrictions on multiple device usage. Most of these restrictions apply to
the MEGA-D.
1. Under Win32, the MEGA-D reserves most of the bus bandwidth, even if the frame rate is
lowered with the SetRate() call. Therefore, for the present it is possible to stream video from
multiple devices simultaneously only if they are on different IEEE 1394 busses, i.e., attached
to separate IEEE 1394 cards. This restriction should be lifted in the near future. There is no
such restriction under Linux.
2. When using multiple devices on the same IEEE 1394 bus, it may be necessary to lower the
frame rate or frame size before starting streaming video. For example, the Dual-DCAM can
use most of the bus bandwidth at 640x480, 15 Hz. Setting the frame rate to 7.5 Hz, or the
frame size to 320x240, will allow more devices to be accessed.
Small Vision System User Manual
C++ Classes
There are three main classes for SVS: classes that encapsulate stereo images, classes that produce the
images from video or file sources, and classes that operate on stereo images to create disparity and 3D
images. These classes are displayed in Figure 5-1. The header file is src/svsclass.h.
The basic idea is to have one class (svsStereoImage) for stereo images and the resultant disparity
images, which performs all necessary storage allocation and insulates the user from having to worry about
these issues. Stereo image objects are produced from video sources, stored image files, or memory buffers
by the svsAcquireImages classes, which are also responsible for rectifying the images according to
parameters produced by the calibration routines. Disparity images and 3D point clouds are produced by
the stereo processing class svsStereoProcess acting on stereo image object, with the results stored
back in the stereo image object. Finally, display classes allow for easy display of the images within a GUI.
Figure 5-2 shows a simple example of using the classes to produce and display stereo disparity results.
The full program example is in samples/stframe.cpp. The basic operations are:
Make a video source object and open it. Which video source is used depends on which
framegrabber interface file has been loaded: see Section 2.1.2.
Make a stereo processing object for producing disparity results from a stereo image.
Make some display window objects for displaying images and disparity results.
Open the video source.
Set the frame size and any other video parameters you wish, and read in rectification
parameters from a file.
Start the video acquisition.
a. Get the next stereo image.
b. Calculate disparity results.
c. Display the results.
Image source
Stereo Image and
Parameter classes
Stereo Processing
Figure 5-1 SVS C++ Classes
Display classes
Small Vision System User Manual
// Make a video source object, using the loaded framegrabber interface
svsVideoImages *videoObject = getVid eoObject();
// Make a stereo processing object
svsStereoProcess *processObject = new svsStereoProcess();
// Open the video source
bool ret = videoObject ->Open();
if (!ret) { …error code… }
// Read in rectification parameters
// Set up display windows
int width = 320, height = 240;
svsWindow *win1 = new svsWindow(width,height);
svsWindow *win2 = new svsWindow(width,height);
// Start up the video stream
videoObject->SetSize(width, height);
ret = videoObject->Start();
if (!ret) { … error code … }
// Acquisition loop
while (1)
// Get next image
svsStereoImage *imageObject = videoObject ->GetImage(400);
if (!imageObject) { … error code …}
// calculate disparity image
// display left image and disparity image
win1->DrawImage(imageObject, svsLEFT);
win2->DrawImage(imageObject, svsDISPARITY);
Figure 5-2 A simple program for video acquisition and stereo processing. The full program is in
Small Vision System User Manual
Parameter Classes
Image frame size and subwindow parameters
Image rectification parameters
Image stereo processing (disparity) parameters
Parameter classes contain information about the format or processing characteristics of stereo image
objects. Each stereo image object contains an instance of each of the above classes. Application programs
can read these parameters to check on the state of processing or the size of images, and can set some of
the parameters, either directly or through class member functions.
Class svsImageParams
Frame size and subwindow parameters for stereo images. In general, the only way these parameters
should be changed is through member functions of the appropriate objects, e.g., using SetSize in the
svsVideoImages class.
Class svsRectParams
Rectification parameters for stereo images. They are used internally by the rectification functions.
Application programs should not change these parameters, and will have few reasons to look at the
parameter values. Rectification parameters are generated initially by the calibration procedure, then
written to and read from parameter files.
Class svsDispParams
Disparity parameters control the operation of stereo processing, by specifying the number of
disparities, whether left/right filtering is on, and so on. Most of these parameters can be modified by
application programs.
Small Vision System User Manual
Stereo Image Class
Stereo image class
The stereo image class encapsulates information and data for a single stereo image pair, along with
any of its processed results, e.g., disparity image or 3D point cloud.
Stereo image objects are usually produced by one of the image acquisition classes
(svsVideoImages or svsFileImages), then processed further by an svsStereoProcess object.
An svsStereoImage object holds information about its own state. For example, there are Boolean
flags to tell if there is a valid set of stereo images, whether they are rectified or not, if a valid disparity
image has been computed, and so on.
The svsStereoImage class handles all necessary allocation of buffer space for images. User
programs can access the image buffers, but should be careful not to de-allocate them or destroy them.
Constructor and Destructor
Constructor and destructor for the class. The constructor initializes most image parameters to default
values, and sets all image data to NULL.
char error[256];
If a member function fails (e.g., if ReadFromFile returns false), then error will usually
contain an error message that can be printed or displayed.
Stereo Images and Parameters
bool haveImages;
// true if we have good stereo images
bool haveColor;
// true if left image color array present
bool haveColorRight;
// true if right image color array present
svsImageParams ip;
// image format, particular to each object
unsigned char *Left(); // left image array
unsigned char *Right(); // right image array
unsigned long *Color(); // left-color image array
unsigned long *ColorRight(); // right-color image array
These members describe the stereo images present in the object. If stereo images are present,
haveImages is true. The stereo images are always monochrome images, 8 bits per pixel.
Additionally, there may be a color image, corresponding to the left image, and a color image for the right
imager, if requested. Color images are in RGBX format (32 bits per pixel, first byte red, second green,
third blue, and fourth undefined). If the left color image is present, haveColor is true. The color
image isn’t used by the stereo algorithms, but can be used in post-processing, for example, in assigning
color values to 3D points for display in an OpenGL window. Similarly, if the right color image is present,
haveColorRight is true. The color images may be input independently of each other.
Frame size parameters for the images are stored in the variable ip. The parameters should be
considered read-only, with one exception: just before calling the SetImage function.
The Left, Right, and Color functions return pointers to the image arrays. User programs should
not delete this array, since it is managed by the stereo object.
Small Vision System User Manual
Rectification Information
bool isRectified; // have we done the rectification already?
bool haveRect;
// true if the rectification params exist
svsRectParams rp; // rectification params, if they exist
The images contained in a stereo image object (left, right and left-color) can be rectified, that is,
corrected for intra-image (lens distortions) and inter-image (spatial offset) imperfections. If the images
are rectified, then the variable isRectified will be true.
Rectification takes place in the svsAcquireImage classes, which can produce rectified images
using the rectification parameters. The rectification parameters can be carried along with the stereo
image object, where they are useful in further processing, for example, in converting disparity images into
a 3D point cloud.
If rectification parameters are present, the haveRect variable is true. The rectification parameters
themselves are in the rp variable.
Disparity Image
bool haveDisparity;
svsDisparityParams dp;
short *Disparity();
// have we calculated the disparity yet?
// disparity image par ameters
// returns the disparity image
The disparity image is computed from the stereo image pair by an svsStereoProcess object. It
is an array of short integers (signed, 16 bits) in the same frame size as the input stereo images. The image
size can be found in the ip variable. It is registered with the left stereo image, so that a disparity pixel at
X,Y of the disparity image corresponds to the X,Y pixel of the left input image. Values of –1 and –2
indicate no disparity information is present: -1 is for low-texture areas, and –2 is for disparities that fail
the left/right check.
If the disparity image has been calculated and is present, then haveDisparity is true. The
parameters used to compute the disparity image (number of disparities, horopter offset, and so on) are in
the parameter variable dp.
The disparity image can be retrieved using the Disparity function. This function returns a pointer
to the disparity array, so it is very efficient. User programs should not delete this array, since it is
managed by the stereo object.
3D Point Array
bool have3D;
int numPoints;
float *X, *Y, *Z;
bool *V;
do we have 3D information?
number of points actually found
3D point arrays
valid 3D point
The 3D point arrays are the 3D points that correspond to each pixel in the left input image. Each
array has the same size (width and height) as the input stereo images. The 3D point array is computed
from the disparity image using the external camera calibration parameters stored in rp. An
svsStereoProcess object must be used to compute it.
Small Vision System User Manual
Each point is represented by a coordinate (X,Y,Z) in a frame centered on the left camera focal point.
The Z dimension is distance from the point perpendicular to the camera plane, and is always positive for
valid disparity values. The X axis is horizontal and the positive direction is to the right of the center of
the image; the Y axis is vertical and the positive direction is down relative to the center of the image (a
right-handed coordinate system). The Boolean array V is used to indicate when there is a valid disparity
reading at a pixel, generating a valid 3D point.
If the 3D array is present, then have3D is true. The actual number of 3D points present in the arrays
is given by numPoints.
File I/O
SaveToFile(char *basename);
ReadFromFile(char *basename);
ReadParams(char *name);
SaveParams(char *name);
saves images and params to files
gets images and params from files
reads just params from file
save just params to file
Images and parameters in a stereo object can be saved to a set of files (SaveToFile), and read back
in from these files (ReadFromFile). The basename is used to create a file set. For example, if the
basename is TESTIMAGE, then the files set is:
left image, if present
right image, if present
left color image, if present
parameter file
Just the parameters can be read from and written to a parameter file, using ReadParams and
SaveParams. These functions take the explicit name of the file, e.g., TESTIMAGE.ini. Parameter
files have the extension .ini, by convention.
Copying Functions
void SetImages(unsigned char *left, // Sets images from user data
unsigned char *right,
unsigned char *color,
unsigned char *color_right,
svsImageParams *ip = NULL,
svsRectParams *rp = NULL,
bool rect = false,
bool copy = f alse);
void CopyFrom(svsStereoImage *si); // copies contents of si to object
These functions are not used in typical applications, since they manipulate the stereo object buffers.
User programs can insert buffer data into a stereo image object using the above functions. These functions
are generally useful for making memory buffers of sequences of images, rather than for initial input of
images. For example, if you want to input images from your own stereo rig, with images stored in
memory, it is recommended to use the svsStoredImages acquisition class, which will produce
svsStereoImage samples. Acquisition classes can perform rectification operations, while the
svsStereoImage class cannot.
Optional parameter information can be supplied with the images, via the parameter arguments;
otherwise, the parameters already present in the object remain the same. If any of the image arguments is
Small Vision System User Manual
NULL, then no image data is inserted for that image. If the copy argument is true, then the buffer
contents are copied onto the stereo image object’s own buffers. If not, then the input buffers are used
Small Vision System User Manual
Acquisition Classes
Base class for all acquisition
Acquire from a video source
Acquire from a file source
Acquire from a memory sourcde
Acquisition classes are used to get stereo image data from video or file sources, and put it into
svsStereoImage structures for further processing. During acquisition, images can be rectified, that
is, put into a standard form with distortions removed. Rectification takes place automatically if the
calibration parameters have been loaded into the acquisition class.
The two subclasses acquire images from different sources. svsVideoImages uses the capture
functions in the loaded svsgrab DLL to acquire images from a video device such as the MEGA-D stereo
head. svsFileImages acquires images from BMP files stored on disk.
Constructor and Destructor
virtual ~svsAcquireImages();
These functions are usually not called by themselves, but are implicitly called by the constructors for
the subclasses svsVideoImages and svsFileImages.
SetRect(bool on);
ReadParams(char *n ame);
SaveParams(char *name);
These functions control the rectification of acquired images. HaveRect() is true when rectification
parameters are present; the normal way to input them is to read them from a file, with ReadParams().
The argument is a file name, usually with the extension .ini. If the acquisition object has rectification
parameters, they can be saved to a file using SaveParams().
Rectification of acquired images is performed automatically if HaveRect() is true, and rectification
processing has been turned on with SetRect(). Calling ReadParams() will also turn on
SetRect(). The state of rectification processing can be queried with GetRect().
If the current image held by the acquisition object is rectified, the IsRect() function will return
Controlling the Image Stream
bool CheckParams()
bool Start()
bool Stop()
svsStereoImage *GetImage(int ms)
Small Vision System User Manual
An acquisition object acquires stereo images and returns them when requested. These functions
control the image streaming process.
CheckParams() determines if the current acquisition parameters are consistent, and returns true if
so. This function is used in video acquisition, to determine if the video device supports the modes that
have been set. If the device is not opened, CheckParams() returns false.
Start() starts the acquisition streaming process. At this point, images are streamed into the object,
and can be retrieved by calling GetImage(). GetImage() will up to ms milliseconds for a new image
before it returns; if no image is available in this time, it returns NULL. If an image is available, it returns
an svsStereoImage object containing the image, rectified if rectification is turned on. The
svsStereoImage object is controlled by the acquisition object, and the user program should not delete
it. The contents of the svsStereoImage object are valid until the next call to GetImage().
Start() returns false if the acquisition process cannot be started. Stop() will stop acquisition.
NOTE: GetImage() returns a pointer to an svsStereoImage object. This object can be
manipulated by the user program until the next call to GetImage(), at which point its data can change.
The user program should not delete the svsStereoImage object, or any of its buffers. All object and
buffer allocation is handled by the acquisition system. If the application needs to keep information around
across calls to GetImage(), then the svsStereoImage object or its buffers can be copied.
Error String
char *Error()
Call this function to get a string describing the latest error on the acquisition object. For example, if
video streaming could not be started, Error() will contain a description of the problem.
Small Vision System User Manual
Video Acquisition
The video acquisition classes are subclasses of svsAcquireImages. The general class is
svsVideoImages, which is referenced by user programs. This class adds parameters and functions
that are particular to controlling a video device, e.g., frame size, color mode, exposure, and so on.
Particular types of framegrabbers and stereo heads have their own subclasses of svsVideoImages. In
general, the user programs won’t be aware of these subclasses, instead treating it as a general
svsVideoImage object.
A particular framegrabber interface class is accessed by copying a DLL file to svsgrab.dll. For
example, for the MEGA-D stereo head and IEEE 1394 interface, copy svspix.dll to svsgrab.dll (see
Section 2.1). Every such interface file defines a subclass of svsVideoImages that connects to a
particular type of framegrabber and its associated stereo head.
To access the svsVideoImages object for the interface file, the special function
svsGetVideoObject() will return an appropriate object.
Video Object
svsVideoImages *svsGetVideoObject()
Returns a video acquisition object suitable for streaming video from a stereo device. The particular
video object that is accessed depends on the video interface file that has been loaded; see Section 2.1, This
function creates a new video object on each call, so several devices can be accessed simultaneously, if the
hardware supports it.
Device Enumeration
int Enumerate()
char **DeviceIDs()
Several MEGA-D and Dual DCAM stereo devices can be multiplexed on the same computer. Any
such device connected to an IEEE 1394 card is available to SVS. The available devices are enumerated by
the Enumerate() function, which returns the number of devices found, and sets up an array of strings
that have the identifiers of the devices. The ID array is returned by the DeviceIDs() function. The
user application should not destroy or write into the array, since it is managed by the video object. The
strings are unique to the device, even when plugged into different machines.
The Enumerate() function rescans the bus each time it is called, so whenever devices are plugged
or unplugged, it can be called to determine which devices are available. DeviceIDs() should only be
called after at least one Enumerate().
MEGA-D and Dual DCAM devicees can be intermixed on the same bus. However, SVS loads only a
single video driver, so any particular SVS application will see only the MEGA-Ds or the Dual DCAMs.
Opening and Closing
bool Open(char *name = NULL)
bool Open(int devnum)
bool Close()
Before images can be input from a stereo device, the device must be opened. The Open() call opens
the device, returning true if the device is available. An optional name can be given to distinguish among
several existing devices. The naming conventions for devices depend on the type of device; typically it is
a serial number or other identifier. These identifiers are returned by the Enumerate() call.
Small Vision System User Manual
Alternatively, a number can be used, giving the device in the order returned by the Enumerate()
function, i.e., 1is the first device, 2 is the second, and so on. A value of 0 indicates any available device.
Upon opening, the device characteristics are set to default values. To set values from a parameter file,
use the ReadParams() function.
A stereo device is closed and released by the Close() call.
Image Framing Parameters
SetSize(int w, int h)
SetSample(int decimation, int binning)
SetRate(int rate)
SetOffset(int ix, int iy, int verge)
SetColor(bool on, bool onr = false)
See Sections 2.1.4, 2.1.5, and 2.1.6 for more information on frame sizes and sampling modes.
These functions control the frame size and sampling mode of the acquired image. SetSize(w,h)
sets the width and height of the image returned by the stereo device. In most cases, this is the full frame
of the image. For example, most analog framegrabbers perform hardware scaling, so that almost any size
image can be requested, and the hardware scales the video information from the imager to fit that size. In
most analog framegrabbers, the sampling parameters (decimation and binning) are not used, and a fullframe image is always returned, at a size given by the SetSize() function.
The digital stereo devices allow the user to specify a desired frame rate. Typically the device defaults
to the fastest frame rate allowed, and the application can choose a different one to minimize bus traffic.
The MEGA-D and Dual-DCAM have different interpretations of the frame rate parameter; see Table
Some stereo devices, such as the MEGA-D, allow the user to specify a subwindow within the image
frame. The subwindow is given by a combination of sampling mode and window size. The sampling
mode can be specified by SetSample(), which sets binning and decimation for the imager. The
MEGA-D supports sub-sampling the image at every 1, 2 or 4 pixels; it also supports binning (averaging)
of 1 or 2 (a 2x2 square of pixels is averaged). For example, with binning =2 and decimation = 2, the full
frame size is 320 x 240 pixels. Using SetSize(), a smaller subwindow can be returned. The offset of
the subwindow within the full frame comes from the SetOffset() function, which specifies the upper
left corner of the subwindow, as well as a vergence between the left and right images.
SetColor() turns color on the left image on or off. Additionally, some applications require color
from the right imager also, and setting the second argument to true will return a color image for the right
imager. Generally, returning color requires more bus bandwidth and processing, so use color only if
0, 1
Normal / 2
Normal / 3
Normal / 4
30 Hz
15 Hz
7.5 Hz
Table 5-1 Frame rates as a function of the SetRate() parameter. MEGA-D frame rates are
determined from the base rate by clock division. Dual-DCAM rates are determined directly
as frames/second.
Small Vision System User Manual
The video frame parameters can be set independently, and not all combinations of values are legal.
The CheckParams() function returns true if the current parameters are consistent.
None of the frame or sampling mode parameters can be changed while images are being acquired,
except for the offset parameters. These can be changed at any time, to pan and tilt the subwindow during
Image Quality Parameters
bool SetExposure(bool auto, int exposure, int gain)
bool SetBalance(bool auto, int red, int blue)
bool SetLevel(bool auto, int brightness, int contrast)
See Section 2.1.9 and 2.1.12 for more information about video quality parameters.
These functions set various video controls for the quality of the image, including color information,
exposure and gain, brightness and contrast. Not all stereo devices support all of the various video modes
described by these parameters.
In general, parameters are normalized to be integers in the range [0,100].
SetExposure() sets the exposure and gain levels for the device. If auto is chosen, the manual
parameters are ignored.
SetBalance() sets the color balance for the device. Manual parameters for red and blue
differential gains are between –40 and 40. If auto is chosen, the manual parameters are ignored.
SetLevel() sets the brightness and contrast for the device. In auto mode, the brightness value is
ignored. Contrast is always set manually.
These functions can be called during video streaming, and their effect is immediate.
Controlling the Video Stream
See the functions in Section 5.5.3.
Small Vision System User Manual
File and Memory Acquisition
The file and memory acquisition classes are subclasses of svsAcquireImages. This classes are
used to input stereo images from files, or from arrays in memory, and present them for processing. Users
who have their own stereo devices, and acquire images into memory, can use these classes to perform
stereo processing with the SVS libraries.
While images are not streamed from files in the same way as from a video source, the function calls
are similar. After opening the file, Start() and then GetImage() function is called to retrieve the
stored image.
A single image file set is repeatedly opened and read in by successive calls to
A file sequence consists of file sets whose basenames end in a 3-digit number. Opening a file set that
is part of a sequence causes the rest of the sequence to be loaded on successive calls to GetImage().
File Image Object
Constructor and destructor for the file acquisition object.
Setting Images from Files
bool Open(char *basename)
bool ReadFromFile(char *basename)
bool Close()
The Open() function opens a file set and reads it into the object; see Section 2.2.2 for information
about file sets. The file set can include calibration parameters that describe the rectification of the images.
Another file can be read into the file object using the ReadFromFile() function. The new images
replace the old ones.
The file is closed with the Close() function.
Stored Image Object
Constructor and destructor for the file acquisition object.
Setting Images from Memory
bool Load(int width, int height,
unsigned char *lim, unsigned char *rim,
unsigned char *cim = NULL, unsigned char *cimr = NULL,
bool rect = false, bool copy = false);
The Load() function sets images from memory into the acquisition object, after which they can be
read out (and optionally rectified) using GetImage(). The left and right monochrome images must be
loaded; the color images are optional. If the images are already rectified, set rect to true.
Normally, user images are passed into the acquisition object as pointers; GetImage() will output
these pointers unless rectification is performed. In this case, the object manages and outputs its own
buffers, which the user should not delete.
Small Vision System User Manual
User images can also be copied into the object’s buffers, if copy is set to true. Here, the object
manages all buffers, and the user program should not destroy them.
Typically, a use program will have images stored in memory, and will want to rectify them as part of
the stereo process. To do so, first load a parameter file into the acquisition object, using
ReadParamFile. Next, turn on rectification by calling DoRect(true). Finally, use Load() to
attach the stored images to the object; calling GetImage() will now return the rectified images.
Small Vision System User Manual
Stereo Processing Classses
Stereo processing class
Multiscale stereo processing class
The stereo processing classses perform stereo processing on stereo images encapsulated in an
svsStereoImage object. The results are stored in the stereo image object. All relevant parameters,
such as calibration information and stereo parameters, are also part of the stereo image object.
The processing class svsStereoProcess handles basic disparity calculation, as well as
conversion of the disparity image into 3D points.
The processing class svsMultiProcess extends stereo processing to perform multiple scale stereo
processing in computing the disparity image. Multiscale processing adds information from stereo
processing at reduced image sizes.
Stereo and 3D Processing
bool CalcStereo(svsStereoImage *si)
bool Calc3D(svsStereoImage *si)
bool CalcPoint3D(int x, int y, svsStereoImage *si,
double *X, double *Y, double *Z)
CalcStereo() calculates a disparity image and stores it in si, assuming si contains a stereo
image pair and its haveDisparity flag is false. To recalculate the stereo results (having set new
stereo processing parameters), set the haveDisparity flag to false and call CalcStereo().
Calc3D() calculates a 3D point array from the disparity image of si. If si does not have a
disparity image, then it is first calculated, and then the point array is computed. The point array is stored
in the stereo image object, and the have3D flag is set.
For some applications, computing the whole 3D array is not necessary; only certain points are needed.
In this case, the function CalcPoint3D() is provided. This function returns true if the disparity at
image point x,y exists, and puts the corresponding 3D values into the X,Y,Z variables. Otherwise, it
returns false and does not change X,Y, or Z.
Multiscale Stereo Processing
bool doIt
This multiscale class computes stereo disparity at the input image resolution, and also at a x2 reduced
image size, then combines the results. Multiscale processing adds additional information, filling in parts
of the disparity image that may be missed at the higher resolution.
The svsMultiProcess class subclasses svsStereoProcess, and is used in exactly the same
way. The boolean variable doIt turns the multiscale processing on or off.
Small Vision System User Manual
Window Drawing Classes
Window class for drawing 2D images
Window class for printing output
The window drawing classes output 2D stereo imagery to the display. The display window relies on
the FLTK cross-platform windowing system (, and provides basic graphical object drawing
in addition to image display. The svsDebugWin class is for text output, useful when debugging
It is also possible to output 3D information, especially point clouds formed from the 3D stereo
reconstruction functions. This display relies on the OpenGL window capabilities of FLTK. For more
information, see the example code in samples/svsglwin.cpp.
Class svsWindow
This class outputs 2D images to a display window. It can output monochrome, color, and false-color
disparity images. In addition, there is an overlay facility for drawing graphical objects superimposed on
the image.
svsWindow objects will downsize the displayed images to fit within their borders, using factors of 2.
For example, a 640x480 image displayed in a 320x200 window will be decimated horizontally by 2 (to
320 columns), and vertically by 4 (to 120 rows). Images smaller than the display window are not
upsampled; they are simply displayed in their normal size in the upper-left corner of the window.
Graphical overlays for the window can be drawn using FLTK drawing functions on the window, e.g.,
lines, circles, etc. svsWindow subclasses the FLTK Fl_Window class.
svsWindow(int x, int y, int w, int h)
Constructor and destructor. The constructor creates a new svsWindow object, displayed at position
x,y of any enclosing FLTK object, and with width w and height h. Typically w and h are multiples of
The window will not be visible until show() is called on it.
DrawImage(svsStereoImage *si, int which = svsL EFT,
void *ovArg = NULL);
Main drawing function. Draws a component of the stereo image object si. The argument which
specifies which component is drawn, according to the following table.
Left and right monochrome images
Left and right color images
Disparity image (displays in green false color)
The optional last argument is passed to any assigned overlay drawing function (see
svsDrawOverlay below).
To clear an image from the window, and reset it to black, use ClearImage().
virtual DrawOverlay(svsImageParams *ip, void *ovArg);
DrawOverlayFn(void (*fn)(svsWindow *, svsImageParams *, void
Small Vision System User Manual
These functions draw overlay information on the image displayed by svsWindow. There are two
ways to draw overlays. One is to subclass svsWindow, overriding the DrawOverlay() function.
Then, the subclass can perform any FLTK drawing within the subclass function.
Another way to use overlays is to assign an overlay function to the svsWindow object, with
DrawOverlayFn. This overlay function is called every time the overlay needs to be drawn.
The last argument to these functions is passed in from the argument specified in the DrawImage()
Class svsDebugWin
This class displays text in a scrollable window.
Constructor and destructor. The x,y arguments specify placement of the upper-left corner of the
window; w and h give the width and height. An optional title (name) can be given.
The window will not be displayed until the show() function is called.
Print(char *str)
Prints a string on the debug window. Each string is printed on a new line, and the window scrolls to
that line.