Towards benchmarking of real-world stereo data

The paper proposes the prediction of stereo matching performance based on analyzing the given stereo data (and not based on test runs of stereo matching algorithms). For justifying our approach we compare results obtained by prediction error analysis (for different stereo matching algorithms) with three different data evaluation techniques: a count of SIFT matches, a mismatch count between census transform features, and the quality of dense optical flow fields based on a total-variation energy minimization. The paper shows that there are reasonable indications that such measures, quantifying matches of features or image regions, correlate with stereo performance to some degree. This study on data evaluation is initiating a new direction of research, and it concludes with the suggestion of studying further measures or more data for the ultimate goal of supporting an adaptive optimization or selection of stereo matching techniques with respect to given image data.