UMVS: Unsupervised Multi-View Stereo for Large-Scale Scenes With Heavy Occlusions

Abstract

Deep learning-based multi-view stereo (MVS) techniques have advanced substantially in recent years. However, because large-scale, annotated datasets are limited, interest has grown in unsupervised MVS methods that do not require ground-truth depth. One primary challenge for these unsupervised approaches is handling occluded regions, as they rely on photometric consistency. To address these issues, we propose Unsupervised Multi-View Stereo for Large-Scale Scenes With Heavy Occlusions (UMVS). Our UMVS comprises three core modules. First, a feature-level perceptual consistency module minimizes reconstruction errors by comparing high-level semantic feature disparities across images. Second, a structured occlusion-generation module improves depth estimation accuracy and completeness by generating augmented samples tailored for contrastive learning. Third, the DLA-Net module serves as a normalization component to overcome the limitations of a narrow receptive field. By aggregating global contextual information, DLA-Net improves depth map estimation accuracy. We evaluated the proposed UMVS on the DTU and Tanks & Temples datasets, and the results demonstrate its effectiveness and superiority.