Optimality in combinations of confidence measures for stereo vision

Abstract

Confidence measures for stereo vision are already popular for some time. Yet, comprehensive results on their performance evaluation are rare. There is still not yet any agreement on answering the question 'what is a good confidence measure'. Very little work has been done for improving discriminativity by exploiting information from combinations of different confidence measures. We present a method to determine an upper bound for performance increase possible by combining given confidence measures. We also provide a solution for a fusion of measures resulting in improved confidence accuracy on popular stereo benchmark data.