Parallel Stereo Matching Based on Edge-Aware Filter

摘要

This paper presents a novel parallel stereo matching algorithm based on edge-aware filter with good performance in accuracy and speed. The initial matching cost is built with census transform and sobel operator. Then the aggregated cost is computed by rolling guidance filter and guided filter. The final disparity is computed by rolling guidance filter and weighted median filter. The key idea is to eliminate the influence of small scale structures when computing weights in aggregation step and post-processing step. The proposed method ranks 17th on Middlebury benchmark and the results cost 52.5ms on one GPU and 33.8ms on two GPUs.