StableMVS: Diffusion-Propelled Stable Multi-View Stereo via Epistemic and Morphological Priors for Urban Depth Ambiguity
This repository contains the official implementation of StableMVS, a robust multi-view stereo framework tailored for urban environments, introducing epistemic priors, morphology-guided propagation, and diffusion-based depth refinement to address low-texture, occlusion, and geometric ambiguity.
- Epistemic Priors: Integrates knowledge from large vision models into cost volume construction.
- Morphological Optimization: Uses structural contours to improve depth smoothness and edge preservation.
- Diffusion-Driven Learning: Enhances illumination-invariant feature learning via perturbation recovery.
git clone https://anonymous.4open.science/r/StableMVS.git
cd StableMVS
conda create -n stablemvs python=3.10
conda activate stablemvs
pip install -r requirements.txtSupported datasets:
python test.py python try.py| Method | MAE (m) | rMAE (%) | RMSE (m) | EAcc (m) | T1 (%) | T3 (%) | T6 (%) | T10 (%) |
|---|---|---|---|---|---|---|---|---|
| CasMVSNet | 0.164 | 0.31 | 0.479 | 0.379 | 58.69 | 91.04 | 96.57 | 97.89 |
| AdaMVS | 0.171 | 0.33 | 0.416 | 0.362 | 52.50 | 89.07 | 96.61 | 98.25 |
| MS-REDNet | 0.199 | 0.37 | 0.518 | 0.425 | 52.26 | 86.38 | 94.83 | 97.20 |
| UCSNet | 0.149 | 0.29 | 0.440 | 0.335 | 59.95 | 92.59 | 97.39 | 98.37 |
| ETNet | 0.195 | 0.37 | 0.474 | 0.400 | 44.21 | 86.04 | 96.42 | 98.02 |
| AggrMVS | 0.153 | -- | -- | -- | -- | -- | 97.00 | 97.96 |
| StableMVS | 0.128 | 0.25 | 0.342 | 0.273 | 62.20 | 94.00 | 98.20 | 98.97 |
Comparison results of different methods on the BlendedMVS dataset. StableMVS ranks first across all metrics.
| Method | EPE | E1 (%) | E3 (%) |
|---|---|---|---|
| MVSNet | 1.49 | 21.98 | 8.32 |
| CasMVSNet | 1.43 | 19.01 | 9.77 |
| CVP-MVSNet | 1.90 | 19.73 | 10.24 |
| Vis-MVSNet | 1.47 | 15.14 | 5.13 |
| EPP-MVSNet | 1.17 | 12.66 | 6.20 |
| UCS-MVSNet | 1.32 | 14.12 | 7.33 |
| TransMVSNet | 0.73 | 8.32 | 3.62 |
| NR-MVSNet | 0.85 | 8.47 | 4.01 |
| UniMVSNet | 0.62 | 9.35 | 3.25 |
| ARAI-MVSNet | 0.67 | 7.91 | 2.95 |
| StableMVS | 0.55 | 7.66 | 2.93 |
More quantitative and qualitative results are available in the paper.
Note: This work is currently under double-blind peer review. Citation will be updated upon publication.
For questions during the review process, please raise an issue in this repository or reach out through the anonymous communication channel provided in the supplementary.