The Stereo Challenge Data Base

Introduction

As previously reported in Fischler [1984] and Hannah [1984], SRI International is implementing a complete, state-of-the-art stereo system that will produce dense three-dimensional (3-D) data from stereo pairs of intensity images.

Ideally, we would assess the capabilities of our system by running it on a data set that has known ground truth against which to compare our results. Unfortunately, such data sets do not currently exist, because of the extremely high cost of the groundwork necessary to measure terrain elevations accurately for a close spacing and to assess the heights of all the temptation and buildings in the area. Lacking such a data set, we can only compare our results against those produced by other stereo systems, or against the perceptions of a human looking at the same and imagery in stereo on a CRT. 

To test our system, currently called STEREOSYS, we have run it on several data sets, including two for which we have results produced by the DIMP stereo system of the US Army engineer topographic laboratories. While comparing our matching results to DiMP results or to human perception of what the correct match should be, we have begun to accumulate a catalog of examples of difficult areas for stereo processing. 

In this report, we describe several data sets that we have processed and discuss the type of problems that our matching algorithms have encountered. This information in part of the “stereo challenge data base” we are assembling to test matching algorithms against; the actual database will contain many more instances of hard-to-match places than are shown in the simple examples examples illustrated here.