A Survey of PatchMatch-Based Multi-View Stereo

Multi-View Stereo (MVS) is a fundamental problem in computer vision, graphics, and machine learning communities, and is a crucial technique for image-based 3D reconstruction. Volume-based Plane-Sweeping (PS) and random-based Patch-Match (PM) are mainstream pipelines for depth-based MVS methods. Although the PS pipeline incorporated with deep networks has attracted generous attention, PM remains the dominant option for real-world applications. This article reviews current PS and PM pipelines and presents in-depth procedures for the latter. Furthermore, we compare and summarize the reconstruction performance of state-of-the-art PS and PM methods on three public benchmarks, providing enlightening observations and inspiring research directions.