MCA-MVSNet: Multi-Scale Context and Attention-Aware Multi-View Stereo Network for Weakly Textured Scene Reconstruction

Abstract

Multi-view stereo (MVS) aims to recover dense 3D structures by exploiting geometric constraints and complementary visual information across multiple views. Although learning-based MVS methods generally achieve better reconstruction performance than traditional approaches, they remain vulnerable to ambiguous local matching cues and unreliable spatial support in weakly textured regions, which may lead to unstable depth estimation and degraded reconstruction completeness. To address these issues, we propose MCA-MVSNet, a multi-scale context and attention-aware multi-view stereo network. Built upon a coarse-to-fine depth estimation framework, MCA-MVSNet integrates a Hierarchical Context-Aware Feature Extractor (HCFE) and an Attention-Aware Depth Estimation (ADE) strategy to jointly enhance feature representation in weakly textured regions and improve the adaptability of spatial support to local feature content. HCFE strengthens multi-scale contextual representation and refines high-resolution local features to provide more reliable features for multi-view matching. ADE employs local-attention-guided offset prediction to make the spatial support used during depth estimation more adaptive to local feature content. Extensive experiments demonstrate that MCA-MVSNet achieves superior reconstruction completeness and overall performance among the compared methods on the DTU dataset, while exhibiting competitive cross-dataset generalization on the Tanks & Temples benchmark.