Deformable Convolution Based Feature Extraction for Multi-View Stereo

Deep learning based multi-view stereo algorithms have shown great reconstruction performance. We propose a method called DCMVSNet for depth inference from multi-view images. We choose to build a cost volume pyramid in a coarse-to- fine manner that leads to a lightweight network. It has an efficient feature extraction module based on deformable convolutional network. The proposed multi-stage network architecture can fully utilize multi-scale feature maps to achieve better reconstruction results. Experiments on popular datasets demonstrate that the proposed method achieves the exceptional reconstruction performance.