Learning Deformable Hypothesis Sampling for Accurate PatchMatch Multi-View Stereo

TL;DR

This paper presents a learnable Deformable Hypothesis Sampler that improves depth estimation accuracy in PatchMatch Multi-View Stereo by learning distribution-sensitive sampling strategies, especially in challenging regions.

Contribution

Introduction of DeformSampler, a learnable module that adapts hypothesis sampling to scene geometry and depth probability distributions, enhancing PatchMatch MVS performance.

Findings

Outperforms state-of-the-art methods on DTU and Tanks  Temples datasets.

Improves depth estimation in discontinuous and weakly-textured regions.

Demonstrates strong generalization capabilities.

Abstract

This paper introduces a learnable Deformable Hypothesis Sampler (DeformSampler) to address the challenging issue of noisy depth estimation for accurate PatchMatch Multi-View Stereo (MVS). We observe that the heuristic depth hypothesis sampling modes employed by PatchMatch MVS solvers are insensitive to (i) the piece-wise smooth distribution of depths across the object surface, and (ii) the implicit multi-modal distribution of depth prediction probabilities along the ray direction on the surface points. Accordingly, we develop DeformSampler to learn distribution-sensitive sample spaces to (i) propagate depths consistent with the scene's geometry across the object surface, and (ii) fit a Laplace Mixture model that approaches the point-wise probabilities distribution of the actual depths along the ray direction. We integrate DeformSampler into a learnable PatchMatch MVS system to enhance depth estimation in challenging areas, such as piece-wise discontinuous surface boundaries and weakly-textured regions. Experimental results on DTU and Tanks \& Temples datasets demonstrate its superior performance and generalization capabilities compared to state-of-the-art competitors. Code is available at https://github.com/Geo-Tell/DS-PMNet.