BE-MVSNet: Boundary- and Edge-Aware Constrained Multi-View Stereo

Learning-based Multi-View Stereo (MVS) methods aim to generate dense point clouds by producing accurate and complete depth maps. However, existing approaches often suffer from insufficient reconstruction completeness in challenging regions such as edges and reflective surfaces, primarily because conventional loss functions fail to provide sufficiently discriminative supervision. To address this issue, we propose BE-MVSNet, an MVS network that incorporates two novel loss functions. Specifically, we design a Boundary-Aware Loss (BAL) that computes the deviation between the ground-truth depth and the predicted depth range, explicitly penalizing pixels whose ground-truth depth falls outside the predicted range, thereby providing more effective supervisory signals for such out-of-range pixels. Additionally, we propose an Edge-Aware Depth Loss (EADL) derived from ground-truth depth maps, which detects geometric edges to provide precise supervision for edge. Extensive experiments demonstrate that BE-MVSNet significantly improves reconstruction completeness in challenging regions while maintaining competitive performance, without introducing additional network complexity.