Efficient multi-view stereo with depth-aware iterations and hybrid loss strategy

Jianfei Jiang, Huimin Ma

Abstract: Highlights•We propose a Depth-Aware Cost Completion (DACC) module, which encodes context-guided depth geometry into the cost volume. DACC contributes to perceiving the geometric shapes of the scene to achieve more accurate depth estimation.•We introduce a Hybrid Loss Strategy (HLS) that enhances the robustness of depth estimation by employing two distinct loss functions at different stages of the process.•We present a novel Depth Representation Consistency (DRC) loss, which seamlessly integrates supervision based on both classification and regression, further enhancing the effectiveness of HLS.•Experimental results indicate that our method achieved state-of-the-art performance with fast inference speed on the DTU dataset and demonstrated powerful generalization ability on the Tanks-and-Temples benchmark, the ETH3D benchmark, and the BlendedMVS dataset.

External IDs:dblp:journals/pr/JiangM26