Effective Co-Segmentation and Data Augmentation for Unsupervised Multi-View Stereo

Recent research has shown that view synthesisbased unsupervised methods may significantly improve multi view stereo (MVS). Although this might not always be the case in reality, current approaches assume that points that correlate across perspectives be coloured similarly. This might lower the quality of the reconstructed signal if the self-supervised data is inaccurate. To solve this problem, we present a paradigm that incorporates semantic co-segmentation along with information augmentation to provide more trustworthy supervision. In particular, we use mutual semantic extraction from many views to direct semantic coherence. By using the forecasting of routine samples as pseudo reality to regularise the preprediction of augmented models, we create an efficient dataaugmentation process that guarantees the robustness of the transformation. Our suggested approaches achieve state-ofthe-art performance on the DTU data set, placing them on par with state-of-the-art supervised algorithms, as presented by our experimental findings. Extensive studies on the Tanks & Temples database further show that the suggested strategy has a high degree of generalizability.