Fast Edge-Aware Cost Aggregation for Stereo Correspondence

Stereo correspondence computes disparity map needed by many high-level computer vision tasks. Recently local stereo approaches show their ability matching global while keeping efficiency. But it still deserves further research for higher performance. In this paper, we propose a fast edge-aware cost aggregation strategy. It constructs 1D support window based on spatial and gradient information of input image, and uses a two-pass scheme to smooth the disparity space image. We also propose an O(1) implementation independent of window size for this scheme. A similar edge-aware median filter is proposed for the post-processing step. According to the evaluation results on the Middlebury stereo dataset, our approach outperforms the state-of-the-art local methods in accuracy and efficiency.