Recurrent Multi-view Stereo Depth Inference with Pyramid of Images

Reconstruction of 3D objects from multiple 2D images is to obtain 3D structural information of spatial objects from 2D images. This paper uses bilinear interpolation to downsample high-resolution input images to reduce memory consumption. A multi-stage depth map estimation method is proposed, which makes full use of the effective information at different scales of the image. The accuracy of the high-resolution depth map is improved by using the low-resolution depth map estimated in the previous stage as the input when estimating the high-resolution depth map. A more efficient regularization method is proposed by fusing Conv, LSTM and U-Net in the regularization stage. We conduct a series of experiments on the DTU dataset and confirm that our network outperforms the original algorithm. Our network improves the reconstruction quality as well as the reconstruction speed.