Cost Aggregation for Stereo Matching Using Total Generalized Variation With Fusion Tensor

Stereo matching methods have achieved remarkable improvements by exploiting various attempts. However, most stereo matching algorithms still suffer from problems such as ambiguous region and inherent ambiguities. In particular, some problems affecting cost aggregation step have the greatest impact on depth results. To resolve the above-mentioned problems, we propose a new cost aggregation method using the modified total generalized variation with fusion tensor. First, two kinds of diffusion tensors are extracted from the guidance color image and the guidance depth map. They are incorporated into an energy functional to obtain the total generalized variation. After formulating the final energy functional, it is optimized via a primal-dual energy minimization method. The performance of the proposed method is experimentally verified by qualitatively and quantitatively comparing the results to those of other algorithms.