Multi-view stereo algorithms based on deep learning: a survey.

Multi-View Stereo (MVS) is a long-standing and fundamental task in computer vision, which aims to reconstruct the 3D geometry of a scene from a set of overlapping images. With known camera parameters, MVS matches pixels across images to compute dense correspondences and recover 3D points. Traditional MVS methods often encounter difficulties in addressing complex scenes due to limitations in robustness. However, recent advancements in deep-learning-based MVS have demonstrated remarkable performance gains, attributed to their enhanced feature extraction and cost volume processing capabilities. This comprehensive survey delves into deep-learning-based MVS algorithms, highlighting key challenges and avenues for future exploration. Our focus is on studies aimed at mitigating algorithmic memory requirements and elevating processing speeds. To better understand the developments in this field, we propose a new taxonomy. The deep-learning-based MVS algorithms are categorized into two categories: scene-oriented and view-oriented methods. Scene-oriented methods typically employ voxels or neural implicit representations as the representative form of the scene and infer the complete scene directly. View-oriented methods focus on estimating the depth map of a single view. Detailed classifications and reviews have been conducted for each category. Furthermore, we introduce several widely used datasets and metrics and empirical evaluation of representative algorithms across three popular datasets. This article concludes by discussing prevalent challenges in the MVS domain and promising research directions for the future. [ABSTRACT FROM AUTHOR]