Probabilistic Mesh Estimation
Differentiable mesh rendering is hindered by the non-differentiability of the triangle selection function. Image softening and hard-edge gradient approximation introduce systematic bias, whereas theoretical compensation methods cost one to two orders of magnitude more than rasterization. We propose probabilistic mesh rendering (PMR), which defines rendering as an expectation over vertex-position distributions. The expectation is estimated by Monte Carlo integration, where each sample is a hard mesh and gradients with respect to the distribution parameters come from the probability density. We instantiate PMR as probabilistic mesh estimation (PME), which starts from a Screened Poisson mesh and jointly optimizes per-vertex Gaussian distributions and spherical-harmonic appearance. Mesh maintenance is performed within the optimization loop, and the mesh is the only object being optimized. On DTU, PME improves PSNR by 6.05 dB and reduces LPIPS by 49% relative to the nvdiffrast controlled baseline, while increasing training time by only 1.22×.
Authors
- Jianhua Cheng (ORCID: https://orcid.org/0000-0002-5204-3550)
- Guojing Zhao (ORCID: https://orcid.org/0009-0001-5267-7457)
- Runze Bai (ORCID: https://orcid.org/0009-0007-3278-905X)
- Guangyan Guo (ORCID: https://orcid.org/0009-0001-7719-5956)
- Zhonghua Liu
- Guanzhen Zhu (ORCID: https://orcid.org/0009-0009-5606-8307)
Institutions
- Harbin Engineering University (CN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24372069
- Primary Topic
- Computer Graphics and Visualization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00