Abstract
Learning-based multi-view stereo (MVS) has gained fine reconstructions on popular datasets. However, supervised learning methods require ground truth for training, which is hard to be collected, especially for the large-scale datasets. Though nowadays unsupervised learning methods have been proposed and have gotten gratifying results, those methods still fail to reconstruct intact results in challenging scenes, such as weakly-textured surfaces, as those methods primarily depend on pixel-wise photometric consistency which is subjected to various illuminations. To alleviate matching ambiguity in those challenging scenes, this paper proposes robust loss functions leveraging constraints beneath multi-view images: 1) Patch-wise photometric consistency loss, which expands the receptive field of the features in multi-view similarity measuring, 2) Robust twoview geometric consistency, which includes a cross-view depth consistency checking with the minimum occlusion. Our unsupervised strategy can be implemented with arbitrary depth estimation frameworks and can be trained with arbitrary large-scale MVS datasets. Experiments show that our method can decrease the matching ambiguity and particularly improve the completeness of weakly-textured reconstruction. Moreover, our method reaches the performance of the state-of-the-art methods on popular benchmarks, like DTU, Tanks and Temples and ETH3D. The code will be released soon.
Abstract (translated)
URL
https://arxiv.org/abs/2203.02156
