Projective Vision Toolkit Homepage

What's New:

The Projective Vision Toolkit (PVT) is a series of utilities available in binary form that allow you to take an image sequence and compute the fundamental matrix and trilinear tensor. The current version only goes as far as computing these two quantities, along with the correspondences that support them. It does so completely automatically, using only natural features. The most important assumption is that the maximum motion of a single feature is limited (usually to 1/3 of the image size). We are able to process images that are more widely separated than those from a video camera, but can not handle very wide separations. If one wishes to perform a reconstruction of the camera positions it is necessary to auto-calibrate (or to a-priori have the calibration). Currently we provide no reconstruction software, but plan to do so in the future. In each of our current examples we have a VRML file (.wrl extension) which shows the reconstruction of the camera positions along with the features that were detected. In this case the reconstruction was obtained by sending the correspondences and calibration information to the Photomodeler package.

The software uses shell scripts to go through the image sequences, and the results can be displayed using a custom viewer. The software is available for Windows (NT,W95,98), and Unix systems (Linux,SGI,Sun). For executing the shell scripts, a Unix system is the easiest to use; you need a shell emulator to execute the shell scripts on Windows. The image sequence viewer runs on Windows and Unix systems and does not require a complex shell script.

For a quickstart first download the software and examples for your particular machine. To view the results extract the examples archive and go to the projexamples2 directory, and then to a particular sub-directory. Execute the displayit shell file for Unix/Linux, or the displayit batch file for Windows. Look at the results for a particular image triple by cycling through up arrow, down arrow; look at different image triples by cycling through left and right arrow. Check the epipolar geometry by clicking on any point (try one in the middle image) with the left mouse button. To actually reproduce the results you can use the doit script, after having put the binaries in the execution path. You should use the command clearPVT in order to delete all the old files before you run doit if you want to be sure of what you are recreating. If you have problems, or want to understand the results then it is necessary to have some grasp of the underlying concepts (see the documentation).

There are fifteen examples, along with their reconstructions in the examples archive. The examples show that the software works on a variety of conditions. All the examples have been processed with the same parameters. They use the default script options, and produced good results. Along with some debugging file output in each directory we also place a VRML file, which contains a graphical view of the reconstruction of the feature points and camera positions.

Current Software Status Download the toolkit (link to forms page) Online Documentation Publications