DCV-MVSNet: Dynamic cost volume for complete multi-view stereo

In Multi-View Stereo (MVS) task, high-quality reconstruction is greatly dependent on robust feature matching and accurate depth estimation. However, previous methods struggle to extract accurate feature representations from challenging areas like poorly textured regions and blurred edges, leading to matching ambiguity. Moreover, the indiscriminate use of the ambiguous static pixel-wise cost volume for regularization further hinders reliable depth estimation. To tackle these challenges, we propose a novel network, termed DCV-MVSNet, to achieve accurate feature extraction and construct a reliable cost volume. In particular, we design an Edge Aware Embedding (EAE) module, implicitly embedding edge and texture information into the multi-scale feature maps for accurate feature matching. Additionally, we propose an Uncertainty-Guided Dynamic Cost Volume (UG-DCV) module to build a complete cost volume for robust depth estimation, by dynamically aggregating the spatial neighboring information guided by uncertainty. In response to the challenge of imbalanced uncertainty data distribution, we introduce an Uncertainty-Driven Attention Loss (UDA Loss). Extensive experiments on the DTU, Tanks & Temples, and ETH3D datasets demonstrate that our DCV-MVSNet can achieve competitive results in terms of both qualitative and quantitative performance compared to other state-of-the-art methods.