Frequency-enhanced representation and cost aggregation for multi-view stereo

Abstract

Cascade-based multi-view stereo (MVS) methods demonstrate exceptional flexibility and efficiency in 3D reconstruction tasks. However, existing methods predominantly focus on pixel-wise correlations in the spatial domain while overlooking the critical role of frequency-domain information essential for modeling challenging scenarios, leading to suboptimal 3D geometric reconstruction. Furthermore, downsampling during multi-scale feature extraction may lead to the loss of critical spatial details, undermining the fidelity of depth estimation in visually degraded scenes. In this paper, we propose FA-MVS, a framework that explicitly incorporates frequency information into multi-scale depth estimation to enhance frequency awareness. Specifically, we propose a frequency-enhanced feature extractor, where frequency representations fused with spatial depth priors are progressively refined to bolster robustness against frequency-sensitive variations. Meanwhile, we propose a frequency-aware cost aggregation module that integrates frequency cues into the cost volume, enabling precise capture of fine details in boundaries and occluded regions. Extensive experiments conducted on the DTU, Tanks and Temples, as well as the challenging ETH3D datasets, demonstrate that our method achieves competitive performance compared to existing advanced approaches while exhibiting strong generalization capability.