EFD-MVSNet: Enhanced Feature Distinctiveness for Multi-View Stereo

Applying the concept of plane-sweeping to deep learning-based Multi-View Stereo (MVS) has seen significant advancements in recent years. However, the challenges such as texture scarcity, weak texture, and reflective scenes can lead to errors in depth estimation, leaving room for improvements in the accuracy and completeness of reconstructed three-dimensional point clouds. In this paper, we delve into the essence of MVS as a one-to-many feature matching task and introduce Enhanced Feature Distinctiveness MVSNet (EFDMVSNet), an innovative depth estimation network. Our focus is on enhancing feature distinctiveness while reducing the computational cost associated with plane-sweeping and improving feature matching accuracy. We propose an efficient Feature Refinement Convolution (FRC) to compress the traditional convolution and enhance its feature contrast. Additionally, we introduce the Aggregated Linear Transformer (ALT) to adapt naturally to the one-to-many MVS task, enabling remote context aggregation within and across images with linear complexity. Our model adopts a coarse-to-fine depth estimation strategy, leveraging accurate estimation results to establish a reliable depth hypothesis range for subsequent stages. To further refine depth estimations, we propose the Depth Completion Embedding (DCE) module, which integrates Uniform Local Binary Pattern (ULBP) operators. Extensive experiments have demonstrated the persuasiveness of our EFD-MVSNet on the DTU and Tanks and Temples (T&T) datasets.