Multiscale Retinex Aggregation to Enable Robust Dense Stereo Correspondence

Stereo correspondence is a traditional but still challenging problem in various computer vision tasks. Although current stereo matching algorithms work well, they are still limited by occlusions, texture less and blurred structures, and particularly illumination differences. By revisiting the cost construction and aggregation step in the stereo correspondence procedure, this paper studies a multiscale retinex aggregation method to achieve accurate dense stereo matching. Our method employs the retinex theory to effectively enhance local contrast and utilize color information to boost the matching cost construction and aggregation. We evaluate our proposed framework on benchmark and surgical stereo data. The experimental results demonstrate that our multiscale retinex aggregation provides a more or comparable accurate dense stereo matching strategy. In particular, our method is robust to heavy illumination differences while giving similar performance to state-of-the-art methods on images with uniform illumination.