Bi-directional Recurrent MVSNet for High-resolution Multi-view Stereo
Taku Fujitomi, Seiya Ito, Naoshi Kaneko, Kazuhiko Sumi · 2021
Learning-based multi-view stereo regularizes cost volumes containing spatial information to reduce noise and improve the quality of a depth map. Cost volume regularization using 3D CNNs consumes a large amount of memory, making it difficult to scale up the network architecture. Recent work proposed a cost-volume regularization method that applies 2D convolutional GRUs and significantly reduces memory consumption. However, this uni-directional recurrent processing has a narrower receptive field than 3D CNNs because the regularized cost at a time step does not contain information about future time steps. In this paper, we propose a cost volume regularization method using bi-directional GRUs that expands the receptive field in the depth direction. In our experiments, our proposed method significantly outperforms the conventional methods in several benchmarks while maintaining low memory consumption.