Cap-CasMVSNet:Capsul-based Cascade Cost Volume for Multi-View Stereo Network

Currently deep learning has been widely used in multi-view stereo, but there is still room for optimization. In this paper, we propose a capsule network-based multi-view stereo network, namely Cap-CasMVSNet. We first introduce a transformer-based filter to highlight the foreground part of the feature. Then aiming at the shortcomings of the fixed receptive field of the traditional convolution kernel, we added a deformable convolution module to the network to enable the convolution to adapt to geometric deformation. We use a capsule neuron to handle global semantic connections between high-level features. Finally, we achieve competitive results on the DTU dataset, showing strong robustness.