Research map / 3D vision / Depth and stereo
Learning Deformable Hypothesis Sampling for Accurate PatchMatch 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 DeformSampler, a learnable plane-indicator and Laplace-mixture probabilistic matcher that makes PatchMatch MVS hypothesis sampling deformable, yielding DS-PMNet with state-of-the-art DTU overall error and competitive Tanks and Temples F1.
Problem
Noisy depth estimation in learning-based PatchMatch multi-view stereo, caused by heuristic, distribution-insensitive depth hypothesis sampling during propagation and perturbation.
Approach
A plane indicator learns intra-view feature self-similarity to produce a plane flow field for structure-aware propagation, and a probabilistic matcher fits a two-component Laplace mixture to per-pixel depth probabilities to drive uncertainty-aware perturbation sampling.
Key results
- On DTU testing set, DS-PMNet reaches overall error 0.290 mm, versus 0.352 mm for PatchMatchNet and 0.295 mm for GeoMVSNet (best among compared methods).
- On Tanks and Temples intermediate set, DS-PMNet fine-tuned on BlendedMVS (Ours*) attains F1 64.16, ranking first; on advanced set F1 is 39.78 (3rd per the text).
- Ablation on DTU: adding both plane indicator and probabilistic matcher lowers overall error from 0.342 mm (baseline) to 0.290 mm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace fixed sampling with the Plane Indicator (ablation) | Learnable PatchMatch solver · DTU testing set | Overall (mm): 0.342 -> 0.322 | Helped |
| Replace fixed sampling with the Probability Matcher (ablation) | Learnable PatchMatch solver · DTU testing set | Overall (mm): 0.342 -> 0.320 | Helped |
| Add both the Plane Indicator and Probability Matcher (ablation) | Learnable PatchMatch solver · DTU testing set | Overall (mm): 0.342 -> 0.290 | Helped |
| Remove the Plane Indicator from the full sampler (ablation) | DS-PMNet · DTU testing set | Overall (mm): 0.290 -> 0.320 | Hurt |
| Remove the Probability Matcher from the full sampler (ablation) | DS-PMNet · DTU testing set | Overall (mm): 0.290 -> 0.322 | Hurt |
| Set ε to -,1,1,1 instead of -,2,2,1 (ablation) | DS-PMNet uncertainty-aware perturbation · DTU testing set | Overall (mm): 0.290 -> 0.312 | Hurt |
| Set ε to -,2,2,2 instead of -,2,2,1 (ablation) | DS-PMNet uncertainty-aware perturbation · DTU testing set | Overall (mm): 0.290 -> 0.312 | Hurt |
| Set ε to -,3,2,1 instead of -,2,2,1 (ablation) | DS-PMNet uncertainty-aware perturbation · DTU testing set | Overall (mm): 0.290 -> 0.300 | Hurt |
| Use DS-PMNet rather than Gipuma | DTU reconstruction · DTU testing set | Acc. (mm): 0.283 -> 0.323 | Hurt |
| Use DS-PMNet rather than GeoMVSNet | DTU reconstruction · DTU testing set | Overall (mm): 0.295 -> 0.290 | Hurt |
| Use DS-PMNet* rather than EPNet | Tanks and Temples reconstruction · Tanks and Temples Intermediate | F1: 63.68 -> 64.16 | Helped |
| Use DS-PMNet* rather than EPNet | Tanks and Temples reconstruction · Tanks and Temples Advanced | F1: 40.52 -> 39.78 | Hurt |
Limitations
- Accuracy on DTU is behind Gipuma, the best method on that metric.
- Tanks and Temples advanced-set F1 trails the best MVSNet variants (EPNet F1 40.52).
Experimental setup
Datasets: DTU BlendedMVS
Benchmarks: DTU Tanks and Temples
Metrics: Accuracy Completeness Overall distance (mm) Precision Recall F1-score
Baselines: Gipuma COLMAP MVSNet CasMVSNet UniMVSNet TransMVSNet PatchMatchNet PatchMatch-RL ACMMP EPNet IS-MVSNet APD-MVS
Builds on: PatchMatchNet PatchMatch-RL MVSNet TransMVSNet
Closest papers by meaning
- Dual-Level Precision Edges Guided Multi-View Stereo with Accurate Planarization (AAAI 2025 · similarity 0.86)
- DVP-MVS: Synergize Depth-Edge and Visibility Prior for Multi-View Stereo (AAAI 2025 · similarity 0.86)
- SD-MVS: Segmentation-Driven Deformation Multi-View Stereo with Spherical Refinement and EM Optimization (AAAI 2024 · similarity 0.85)
- SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation (CVPR 2025 · similarity 0.85)
- PatchFusion: An End-to-End Tile-Based Framework for High-Resolution Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.85)
- DEFOM-Stereo: Depth Foundation Model Based Stereo Matching (CVPR 2025 · similarity 0.85)
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.