1. The input images should be GREY SCALE (or 8 bits) images, i.e. NOT colour images.
2. The input images should be epipolar images, or rectified images. In the rectified images, the matching points should lie on the same horizontal scanlines on the left and right images.
3. This fast stereo matching algorithm produces dense disparity map for every pixels in the image. The algorithm uses a pyramid structure, fast correlation, rectangular subregioning and dynamic programming techniques.
4. The images sizes of the left and the right images should be the same. The actual numbers of rows and columes can be any number, i.e. they do not have to be the power of two. The images do not have to be squared. For the demo version, the size of the input images cannot be larger than 640x512 in order to reduce the computation load on our server. If an image size is over 640x512, "Format conversion failure (ConvertFromPnm)" error will occur.
5. Most of the common image formats are readable.
6. The disparity search range is specified in the top level of the pyramid. For example, if the maximum disparity range in the original image is 25 pixels and you specify the level of pyramid as 3, then the disparity search range to be used for the demo should be roughly 6 pixels (25 divide by 2, then divide by 2).
7. The obtained disparity map has been normalised so that the disparity values lie within [0, 255].
8. When you use your own images, make sure they are readable/displayable through the Internet. One easy test to see whether your images can be read by the demo is to display the images yourself using a web browser. For example, try to access your image by giving the URL location "http://extra.cmis.csiro.au/IA/changs/stereo/ballL.gif". If the images can be displayed inside the browser, your image should appear within the browser. Otherwise "xv"(on UNIX) may be called to display the image. If your image is not readable or directory path is not correct, an error message, such as "File Not found" may appear.
9. Other related papers are: "A Fast Stereo Matching Method (DICTA'97)", "Multi-Resolution Stereo Matching Using Maximum-Surface Techniques (DICTA'99)" (A longer version of the paper is available here), and "Rectangular Subregioning and 3D Maximum-Surface Techniques for Fast Stereo Matching (IEEE SMBV'01)".
10. The time that the demo spent on processing your images depends on a number of factors. These include the image size, disparity search range, the complexity of the image or object itself, the network traffic, the current work load of our server etc. The machine type of our local server is 360MHz SUN Ultra 10 running Solaris 2.7.
11. If you have any comments/suggestions/questions, please email me.
12. The default images used ("baseball on newspaper") for the demo is from W. Hoff of the University of Illinois.
13. Other image pairs that you can try are: corr, pentagon, coal.
14. We thank Norrie Bennie and Kevin Cheong for their help in setting up this web demo.

Please send your comments to me! © 1998 by Changming Sun - all rights reserved.