Multi-view Stereo Combined with Space Propagation and Pixel-Level Refinement

This paper proposes a new general framework for multi-view dense reconstruction. We divide the dense reconstruction into three parts: depth map extraction, optimization and fusion, and we proposed some new ideas in each part. We extract depth maps by introducing space propagation, then we iterate between pixel-level updating and filtering for depth map optimization. At last, depth map fusion is executed in a highly parallel way. Our algorithm has been validated on various benchmarks and crowded scene data, and has achieved very good results in terms of efficiency, accuracy and completeness.