Research map / 3D vision / Depth and stereo
DVP-MVS: Synergize Depth-Edge and Visibility Prior for Multi-View Stereo
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DVP-MVS, a traditional PatchMatch multi-view stereo method that adds depth-edge aligned priors (Depth Anything V2 + Roberts operator + erosion-dilation) and cross-view visibility priors to stabilize patch deformation, plus geometry-driven propagation and refinement.
Problem
Patch deformation-based multi-view stereo is unstable in textureless areas because deformed patches skip depth edges and cover occluded or invisible regions, causing estimation deviation.
Approach
Align coarse monocular depth (Depth Anything V2) with Roberts-operator edges via an erosion-dilation strategy to get homogeneous boundaries that constrain anchor selection, restore visibility maps with cross-view depth reprojection, and use visible hemispherical normal aggregation and epipolar-line depth intervals for propagation and refinement.
Key results
- On ETH3D test at 2cm, DVP-MVS reaches F1 89.60 vs 87.44 for the APD-MVS baseline, with completeness 89.41.
- On Tanks & Temples intermediate, F1 64.76 vs 63.64 for APD-MVS, the highest completeness of 78.69 recall.
- Ablation: removing the whole depth-edge aligned prior gives the largest F1 drop (86.53 vs 88.67 at 2cm).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DVP-MVS instead of HPM-MVS++ | ETH3D train · ETH3D train, 2cm | F1: 87.09 -> 88.67 | Helped |
| use DVP-MVS instead of HPM-MVS++ | ETH3D test · ETH3D test, 2cm | F1: 89.02 -> 89.60 | Helped |
| use DVP-MVS instead of EPNet | Tanks & Temples intermediate · Tanks & Temples intermediate | F1: 63.68 -> 64.76 | Helped |
| use DVP-MVS instead of EPNet | Tanks & Temples advanced · Tanks & Temples advanced | F1: 40.52 -> 40.23 | Hurt |
| add the depth-edge aligned prior (ablation) | DVP-MVS · ETH3D, 2cm | F1: 86.53 -> 88.67 | Helped |
| add region erosion (ablation) | DVP-MVS · ETH3D, 2cm | F1: 88.02 -> 88.67 | Helped |
| add region dilation (ablation) | DVP-MVS · ETH3D, 2cm | F1: 87.56 -> 88.67 | Helped |
| add pixel-wise filtering (ablation) | DVP-MVS · ETH3D, 2cm | F1: 87.75 -> 88.67 | Helped |
| add the cross-view prior (ablation) | DVP-MVS · ETH3D, 2cm | F1: 86.76 -> 88.67 | Helped |
| add visibility map restoration (ablation) | DVP-MVS · ETH3D, 2cm | F1: 87.84 -> 88.67 | Helped |
| add visibility-aware patch deformation (ablation) | DVP-MVS · ETH3D, 2cm | F1: 87.39 -> 88.67 | Helped |
| add the geometry-driven module (ablation) | DVP-MVS · ETH3D, 2cm | F1: 86.92 -> 88.67 | Helped |
Limitations
- Relies on monocular depth and Roberts edge quality; depth-edge alignment depends on fixed thresholds.
- Accuracy is not best on several metrics (e.g., TNT advanced precision lower than baselines).
Experimental setup
Datasets: ETH3D Tanks & Temples
Benchmarks: ETH3D Tanks & Temples
Metrics: F1 score accuracy completeness precision recall
Baselines: PatchMatchNet IterMVS-LS MVSTER EPP-MVSNet EPNet TAPA-MVS PCF-MVS ACMM ACMP ACMMP SD-MVS APD-MVS
Builds on: APD-MVS ACMMP Depth Anything V2 PatchMatch
Closest papers by meaning
- MSP-MVS: Multi-Granularity Segmentation Prior Guided Multi-View Stereo (AAAI 2025 · similarity 0.91)
- Dual-Level Precision Edges Guided Multi-View Stereo with Accurate Planarization (AAAI 2025 · similarity 0.91)
- SD-MVS: Segmentation-Driven Deformation Multi-View Stereo with Spherical Refinement and EM Optimization (AAAI 2024 · similarity 0.89)
- Planar Prior Assisted PatchMatch Multi-View Stereo (AAAI 2020 · similarity 0.89)
- PatchFusion: An End-to-End Tile-Based Framework for High-Resolution Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.86)
- Learning Deformable Hypothesis Sampling for Accurate PatchMatch Multi-View Stereo (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.