Research map / 3D vision / Depth and stereo
DeMatch: Deep Decomposition of Motion Field for Two-View Correspondence Learning
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 DeMatch, a network that decomposes a contaminated two-view motion field into learnable motion-pattern-driven smooth sub-fields and recovers a cleaner field for inlier classification, achieving implicit regularization with lower cost.
Problem
Two-view correspondence learning (outlier rejection) needs motion-field coherence and smoothness, but explicit regularization is computationally heavy and local smoothing over-smooths discontinuities caused by large scene disparities.
Approach
Inspired by Fourier expansion, learnable motion-pattern tokens act as a basis; putative motion vectors are softly clustered by attention into a few smooth sub-fields, which are then used to recover a cleaner motion field. A masked decomposition removes likely outliers using the previous layer's inlier probabilities, yielding implicit regularization and piecewise smoothness.
Key results
- On YFCC100M relative pose estimation with the weighted eight-point algorithm, DeMatch reaches AUC@5° 30.89 vs ConvMatch 26.83 and LMCNet 22.35.
- On SUN3D with weighted eight-point, DeMatch AUC@10° is 23.10 vs ConvMatch 22.23.
- DeMatch runs in 26.36 ms with 2.346 GFLOPs and 5.853M parameters, far faster than LMCNet (227.15 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeMatch instead of NCMNet | Two-view correspondence learning · YFCC100M | AUC@5° with weighted eight-point algorithm: 26.89 -> 30.89 | Helped |
| Use DeMatch instead of NCMNet | Two-view correspondence learning · SUN3D | AUC@10° with weighted eight-point algorithm: 16.84 -> 23.10 | Helped |
| Use DeMatch instead of ConvMatch | Visual localization · Aachen Day-Night Day | Localization accuracy at (0.25m,2°): 84.5 -> 85.2 | Helped |
| Use DeMatch instead of ConvMatch | Visual localization · Aachen Day-Night Night | Localization accuracy at (5.0m, 10°): 91.8 -> 94.9 | Helped |
| Add masked decomposition (ablation) | DeMatch · YFCC100M | AUC@10° with weighted eight-point algorithm: 51.10 -> 52.67 | Helped |
| Add global information enhancement (ablation) | DeMatch · YFCC100M | AUC@10° with weighted eight-point algorithm: 46.55 -> 52.67 | Helped |
| Add the decomposition process (ablation) | DeMatch · YFCC100M | AUC@10° with weighted eight-point algorithm: 43.64 -> 52.67 | Helped |
| Add decomposition and global information enhancement (ablation) | DeMatch · YFCC100M | AUC@10° with weighted eight-point algorithm: 38.48 -> 52.67 | Helped |
| Use standard attention instead of vanilla attention (ablation) | DeMatch · YFCC100M | AUC@10° with weighted eight-point algorithm: 50.93 -> 52.67 | Helped |
| Increase coherence constraint layers from L=3 to L=5 (ablation) | DeMatch · YFCC100M | AUC@10° with weighted eight-point algorithm: 45.38 -> 52.67 | Helped |
| Increase motion patterns from K=32 to K=48 (ablation) | DeMatch · YFCC100M | AUC@10° with weighted eight-point algorithm: 49.85 -> 52.67 | Helped |
| Initialize basis with uniform rather than normal distribution (ablation) | DeMatch · YFCC100M | AUC@5° with RANSAC: 32.98 -> 33.73 | Helped |
Surprising findings
- Learnable basis initialization barely matters: all methods except all-zeros and all-ones give similar results.
- LMCNet's explicit regularization is much slower (227 ms) than DeMatch's implicit decomposition.
Limitations
- Computational cost is higher than some lightweight baselines such as PointCN and OANet in time and parameters.
- Gains over ConvMatch on SUN3D are small under RANSAC (e.g., 18.66 vs 18.45 at AUC@10°).
Experimental setup
Datasets: YFCC100M SUN3D
Benchmarks: YFCC100M SUN3D Aachen Day-Night
Metrics: AUC@5° AUC@10° AUC@20° Time (ms) Flops (G) Param (M) Mem (MB) visual localization accuracy at (0.25m,2°), (0.5m,5°), (5m,1
Baselines: NN GMS LPM CRC VFC PointCN OANet CLNet MS2DGNet NCMNet LMCNet ConvMatch
Builds on: OANet ConvMatch LMCNet PointCN graph attention network
Closest papers by meaning
- DeMo: Deep Motion Field Consensus with Learnable Kernels for Two-view Correspondence Learning (AAAI 2025 · similarity 0.90)
- Frequency Decoupling for Motion Magnification via Multi-Level Isomorphic Architecture (CVPR 2024 · similarity 0.84)
- BCLNet: Bilateral Consensus Learning for Two-View Correspondence Pruning (AAAI 2024 · similarity 0.84)
- DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video (CVPR 2024 · similarity 0.84)
- Decompositional Neural Scene Reconstruction with Generative Diffusion Prior (CVPR 2025 · similarity 0.84)
- Degrees of Freedom Matter: Inferring Dynamics from Point Trajectories (CVPR 2024 · similarity 0.83)
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.