MLUF-MVSNet: an improved feature extraction and enhanced attention for unsupervised contrastive learning in multi-view 3D reconstruction

This is the authors' abstract. We don't add key points for this paper.

Unsupervised multi-view stereo (MVS) techniques seek to reconstruct three-dimensional scenes without dependence on ground truth annotations. Nevertheless, current methods often face challenges in capturing long-range dependencies and global contextual cues, which adversely affects reconstruction accuracy in regions characterized by weak or repetitive textures. In this study, we introduce MLUF-MVSNet, an unsupervised contrastive learning framework for multi-view 3D reconstruction that enhances feature extraction and attention mechanisms to achieve precise and comprehensive reconstructions in such challenging areas. Specifically, we propose a selective structured state-space model-based feature pyramid network (FPN) to more effectively capture global contextual information. Additionally, we integrate a multi-scale convolutional approach combined with a cross-dimensional information aggregation attention module to mitigate the loss of fine-grained details. Extensive evaluations on the DTU and Tanks & Temples benchmark datasets demonstrate that our unsupervised MVS framework exhibits competitive performance compared to most state-of-the-art (SOTA) methods.