Research map / 3D vision / Depth and stereo
DKM: Dense Kernelized Feature Matching for Geometry Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the correlation-volume regressor with kernelized regression and a cosine embedding decoder (ablation) | DKM global matcher · MegaDepth-1500 (global matcher ablation) | AUC @5°: 57.0 -> 58.1 | Helped |
| Use linear rather than cosine coordinate embeddings (ablation) | DKM global matcher · MegaDepth-1500 (global matcher ablation) | AUC @5°: 58.1 -> 57.9 | Hurt |
| Use stacked feature maps and the proposed input representation rather than baseline inputs (ablation) | DKM warp refiner · MegaDepth-1500 (warp refiner ablation) | AUC @5°: 56.5 -> 58.1 | Helped |
| Use depthwise warp refiners rather than baseline refiners (ablation) | DKM warp refiner · MegaDepth-1500 (warp refiner ablation) | AUC @5°: 54.9 -> 58.1 | Helped |
| Add certainty-weighted match sampling (ablation) | DKM match sampling · Two-view pose estimation (sampling ablation) | AUC @5°: 42.9 -> 56.1 | Helped |
| Balance certainty-sampled matches using reciprocal KDE density (ablation) | DKM match sampling · Two-view pose estimation (sampling ablation) | AUC @5°: 56.1 -> 58.1 | Helped |
| Increase training resolution from 384×512 to 480×640 (ablation) | DKM · Two-view pose estimation (resolution ablation) | AUC @5°: 58.1 -> 58.9 | Helped |
| Increase training resolution from 480×640 to 540×720 (ablation) | DKM · Two-view pose estimation (resolution ablation) | AUC @5°: 58.9 -> 59.4 | Helped |
| Concatenate reverse-warp matches with forward-warp matches (ablation) | DKM · Two-view pose estimation (bidirectionality ablation) | AUC @5°: 59.4 -> 60.4 | Helped |
| Make the kernel temperature learnable (ablation) | DKM global matcher · Global matching | performance | No clear effect |
| Use a squared exponential kernel instead of the exponential cosine similarity kernel (ablation) | DKM global matcher · Global matching | performance | No clear effect |
Closest papers by meaning
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · similarity 0.88)
- EDM: Equirectangular Projection-Oriented Dense Kernelized Feature Matching (CVPR 2025 · similarity 0.87)
- RoMa: Robust Dense Feature Matching (CVPR 2024 · similarity 0.86)
- Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · similarity 0.86)
- Dense Match Summarization for Faster Two-view Estimation (CVPR 2025 · similarity 0.86)
- RUBIK: A Structured Benchmark for Image Matching across Geometric Challenges (CVPR 2025 · 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.