3D Reconstruction Based on Cyclic Multi-View Stereo Network

Currently, the multi-view Stereo algorithm performs well in the neighborhood reconstruction, but the MVSNet method still has a huge memory consumption problem in the reconstruction of large scenes. This paper presents a recursive neural network based on GRU (gated recurrent unit) to reduce memory consumption. Firstly, the feature map of the reference image and the neighborhood image is extracted, and then the feature map of the neighborhood image is mapped to the cone of the reference image through the homography transformation. Then we use GRU to regularize the feature graphs at different depths, and finally filter and refine the depth features to further improve the quality of 3d point clouds. Compared with Convolutional Neural Networks, the Convolutional GRU regularization greatly reduces the memory consumption and improves the scalability. Experimental results show that the proposed method can reduce memory consumption and achieve high quality modeling under high resolution scenarios.