Research map / 3D vision / Depth and stereo
DeMo: Deep Motion Field Consensus with Learnable Kernels 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 DeMo, a deep motion-field consensus network for outlier rejection in two-view correspondence learning that uses RKHS interpolation with learnable Gaussian kernels, scene-adaptive sampling, and a local consensus complementation block.
Problem
Rejecting outliers among putative correspondences between two images by capturing global motion consensus, which existing learning methods largely neglect, while keeping computation manageable for the cubic cost of explicit interpolation.
Approach
Formulates motion-field consensus as Tikhonov-regularized RKHS interpolation with learnable deep Gaussian kernels (closed-form solution), reduces its O(N^3) cost with a soft scene-adaptive sampling of M motions, and adds a local consensus complementation block over feature and spatial channels to handle depth discontinuities.
Key results
- On YFCC100M, DeMo reaches AUC@10 of 54.82 (weighted eight-point) versus 52.17 for UMatch and 53.32 for DeMatch.
- On HPatches with DLT, DeMo gets 53.79% accuracy at 3px versus 47.07% for UMatch and 46.90% for DeMatch.
- Ablation shows removing the DMFC block drops YFCC100M AUC@10 from 54.82 to 27.21.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DeMo instead of DeMatch | two-view correspondence learning · YFCC100M, weighted eight-point, AUC@10° | AUC@10°: 52.67 / 52.37 -> 54.82 / 54.42 | Helped |
| use DeMo instead of DeMatch | two-view correspondence learning · SUN3D, weighted eight-point, AUC@10° | AUC@10°: 23.10 / 18.66 -> 23.18 / 18.85 | Helped |
| use DeMo instead of DeMatch | homography estimation · HPatches, Acc.@5px | accuracy: 60.52 / 82.76 -> 63.79 / 85.52 | Helped |
| use DeMo instead of DeMatch | visual localization · Aachen Day-Night, night, (0.25m, 2°) / (0.5m, 5°) / (5.0m, 10°) | localized queries (%): 73.5 / 84.7 / 94.9 -> 74.5 / 84.7 / 94.9 | No clear effect |
| add the LCC block (ablation) | DeMo · YFCC100M | AUC@10°: 49.61 -> 54.82 | Helped |
| add spatial-channel processing (ablation) | DeMo · YFCC100M | AUC@10°: 53.78 -> 54.82 | Helped |
| add the DMFC block (ablation) | DeMo · YFCC100M | AUC@10°: 27.21 -> 54.82 | Helped |
| add DMFC interpolation (ablation) | DeMo · YFCC100M | AUC@10°: 53.65 -> 54.82 | Helped |
| use scene-adaptive sampling instead of probability-based selection (ablation) | DeMo · YFCC100M | AUC@10°: 52.15 -> 54.82 | Helped |
| use scene-adaptive sampling instead of random selection (ablation) | DeMo · YFCC100M | AUC@10°: 52.28 -> 54.82 | Helped |
Surprising findings
- On SUN3D with RANSAC, DeMo's AUC@5 (7.48) is not clearly ahead of some baselines, and on HPatches with RANSAC it is lower than LMCNet and others at some thresholds.
Limitations
- Computation time is higher than some baselines (64.46 ms vs 9.33 ms for PointCN on HPatches).
- Gains over strong baselines on several RANSAC-based metrics are small.
Experimental setup
Datasets: YFCC100M SUN3D HPatches Aachen Day-Night
Benchmarks: YFCC100M SUN3D HPatches Aachen Day-Night
Metrics: AUC@5/10/20 homography accuracy at 3/5/10 px F-score runtime localization accuracy at day/night thresholds
Baselines: CRC GMS LPM VFC PointCN OANet LMCNet NCMNet ConvMatch UMatch DeMatch
Builds on: OANet ConvMatch DeMatch ContextNorm HLoc
Closest papers by meaning
- DeMatch: Deep Decomposition of Motion Field for Two-View Correspondence Learning (CVPR 2024 · similarity 0.90)
- BCLNet: Bilateral Consensus Learning for Two-View Correspondence Pruning (AAAI 2024 · similarity 0.87)
- VGGSfM: Visual Geometry Grounded Deep Structure From Motion (CVPR 2024 · similarity 0.84)
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · similarity 0.84)
- ColabSfM: Collaborative Structure-from-Motion by Point Cloud Registration (CVPR 2025 · similarity 0.84)
- Motion2VecSets: 4D Latent Vector Set Diffusion for Non-rigid Shape Reconstruction and Tracking (CVPR 2024 · similarity 0.84)
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.