.png){kind=icon}
.png){kind=icon}
Return
Rethinking probability volume for multi-view stereo: A probability analysis method
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-025-06284-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The existing learning-based multi-view stereo (MVS) models primarily focus on predicting depth maps through a cascaded structure to achieve more robust reconstruction results. However, they often emphasize improving the quality of stereo matching while overlooking the importance of depth hypotheses. In this paper, we propose a novel MVS model from the perspective of probability volume analysis. First, the guiding effect of the probability volume is considered for depth refinement. Ideally, the probability distribution along the depth dimension of the probability volume follows an unimodal pattern. We design an unimodal curve to fit this pattern. Then, a reasonable depth refinement range is adaptively selected for each pixel position based on a predefined probability threshold. Additionally, considering that matching noise may cause the probability volume to appear as a blurred unimodal peak, we design the probability volume split-merge module (PVS-PVM). This module performs a peak search based on conditional constraints, splitting the probability volume into main and sub probability volumes, then computes the two sets of depth hypotheses from them. Finally, the new main and sub probability volumes are computed based on these depth hypotheses and merged to predict the depth. This approach allows for a more comprehensive consideration of the regions with higher probability, improving the robustness of depth hypotheses. Experimental results demonstrate that our method effectively utilizes probability volume information to guide depth map refinement and yields enhanced reconstruction results on the DTU and Tanks & Temples datasets. Our code will be released at https://github.com/zongh5a/ProbMVSNet.
Keywords:
Multi-view stereo
Adaptive depth hypotheses
Depth estimation
Deep learning
Point cloud
Journal
IF:
3.5
Papers:
7.5K
Citations:
1.7W
Organization
No organization information available