DAR-MVSNet: a novel dual attention residual network for multi-view stereo

Learning-based multi-view stereo (MVS) has shown great promise in the field of 3D reconstruction. However, existing MVS methods suffer from fixed receptive field sizes during feature learning. This issue leads to information loss and affects the understanding of the geometric structure of the scene, posing a challenge to the reconstruction quality of regions with complex geometric structures and lighting conditions. Therefore, we propose DAR-MVSNet, which consists of a dual-attention-guided feature pyramid network (DA-FPN) and a 3D residual U-net module (3D-RUM). DA-FPN includes two modules: attention-based context extraction module (ACEM) and self-attention-based module (SAM). ACEM is proposed to dilate the receptive field and preliminarily filter deep features through multiple dilated convolutions, spatial and channel attention. To further eliminate redundant characteristics, SAM is proposed to enhance the representation capability of depth features. Moreover, 3D-RUM is designed to enhance feature transfer and information flow, thereby addressing the problem of severe global feature information loss. This article demonstrates the effectiveness of DAR-MVSNet through an extensive series of experiments. The result on the DTU dataset and Tanks and Temples benchmark in comparison to state-of-the-art MVS methods verify its superior performance.