ML-MVSNet: Multi-View Stereo with Local Awareness and Geometric Guidance

Learning-based multi-view stereo (MVS) methods have achieved remarkable progress in recent years. However, many existing approaches neglect the role of neighborhood information in feature point matching and lack explicit geometric guidance in depth prediction, which may lead to suboptimal reconstruction quality. To address these challenges, we propose ML-MVSNet, a novel MVS network that incorporates local information awareness and geometric guidance. Specifically, a neighborhood information awareness strategy is introduced during feature matching to alleviate the limitations of insufficient matching. Additionally, we integrate geometric guidance via iterative cost volume regularization. Finally, we employ a dualbranch subnetwork to refine and upsample depth maps, further improving reconstruction quality. Extensive experiments on the DTU dataset and Tanks and Temples benchmark demonstrate that ML-MVSNet achieves competitive performance compared to state-of-the-art methods.