Research map / 3D vision / Depth and stereo
Dual-Level Precision Edges Guided Multi-View Stereo with Accurate Planarization
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 DPE-MVS, a PatchMatch multi-view stereo method that uses fine (Canny) and coarse (Roberts+Hough) edges to build more accurate plane models, guide sampling, and adaptively resize patches for low-textured and stochastic areas.
Problem
Multi-view stereo reconstruction of low-textured and stochastic-textured areas, where plane-model construction in PatchMatch methods suffers from crossing object boundaries and limited perception range.
Approach
Extends APD-MVS with dual-level edges (fine Canny edges and coarse Roberts/Hough segmentation edges) to constrain RANSAC plane fitting, expand the anchor search range, guide non-local hypothesis sampling, and adaptively adjust the center patch size (enlarge in low-textured areas, discard in stochastic areas).
Key results
- On ETH3D test at 2cm, DPE-MVS reaches F1 89.48 with completeness 88.77, the best among compared methods (APD-MVS 87.44).
- On Tanks & Temples Intermediate, DPE-MVS achieves recall 78.16 (best) and F1 63.98; on Advanced F1 40.20 with recall 56.45 (best recall).
- Ablation on ETH3D training at 2cm: full DPE-MVS improves F1 from 75.80 (baseline) to 78.11 at 1cm and from 94.37 to 95.53 at 5cm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Edge-Guided Non-Local Sampling (ES) (ablation) | APD-MVS · ETH3D training | F1: 75.80 -> 76.67 (1cm); 94.37 -> 94.38 (5cm) | Helped |
| Add Perception Range Expansion (PE) (ablation) | APD-MVS · ETH3D training | F1: 75.80 -> 76.27 (1cm); 94.37 -> 94.54 (5cm) | Helped |
| Add Plane Construction Optimization (PO) to ES and PE (ablation) | DPE-MVS · ETH3D training | F1: 77.13 -> 77.79 (1cm); 94.81 -> 95.05 (5cm) | Helped |
| Add Adaptive Patch Size Adjustment (AA) to ES and PE (ablation) | DPE-MVS · ETH3D training | F1: 77.13 -> 77.26 (1cm); 94.81 -> 95.26 (5cm) | Mixed |
| Add ES to ACMM (ablation) | ACMM · ETH3D training | F1: 67.58 -> 69.34 (1cm); 78.86 -> 80.47 (2cm); 87.68 -> 89.17 (5cm); 91.70 -> 92.83 (10cm); 94.41 -> 95.07 (20cm) | Mixed |
| Add ES to ACMP (ablation) | ACMP · ETH3D training | F1: 68.72 -> 72.45 (1cm); 79.79 -> 83.01 (2cm); 88.32 -> 90.70 (5cm); 92.03 -> 93.72 (10cm); 94.43 -> 95.53 (20cm) | Helped |
| Add ES to ACMMP (ablation) | ACMMP · ETH3D training | F1: 71.57 -> 76.01 (1cm); 83.42 -> 86.60 (2cm); 92.03 -> 94.07 (5cm); 95.54 -> 97.07 (10cm); 97.37 -> 98.37 (20cm) | Helped |
Surprising findings
- Adaptive patch discarding in stochastic areas accelerates deformable PM, saving 5.72 s per iteration despite added edge processing.
- Ablation of ES in ACMM shows better 1cm/2cm F1 for NESP than ES, attributed to ACMM lacking plane priors.
Limitations
- Lower precision on Tanks & Temples Advanced set due to extensive low-textured areas, prioritizing recall.
- Weaker at fine scene details where learning-based methods excel.
- Adds up to 2GB extra memory and parameters (alpha, beta, sigma) are empirically set.
Experimental setup
Datasets: ETH3D Tanks & Temples
Benchmarks: ETH3D Tanks & Temples
Metrics: accuracy completeness F1-score precision recall
Baselines: APD-MVS TSAR-MVS SD-MVS HPM-MVS ACMM ACMMP ACMP PatchMatchNet IterMVS EPNet GoMVS DS-PMNet
Builds on: APD-MVS ACMH TSAR-MVS ACMM ACMMP
Closest papers by meaning
- Planar Prior Assisted PatchMatch Multi-View Stereo (AAAI 2020 · similarity 0.92)
- DVP-MVS: Synergize Depth-Edge and Visibility Prior for Multi-View Stereo (AAAI 2025 · similarity 0.91)
- MSP-MVS: Multi-Granularity Segmentation Prior Guided Multi-View Stereo (AAAI 2025 · similarity 0.90)
- Mesh-Guided Multi-View Stereo With Pyramid Architecture (CVPR 2020 · similarity 0.90)
- SD-MVS: Segmentation-Driven Deformation Multi-View Stereo with Spherical Refinement and EM Optimization (AAAI 2024 · similarity 0.88)
- MonSter: Marry Monodepth to Stereo Unleashes Power (CVPR 2025 · similarity 0.87)
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.