Research map / 3D vision / Human pose estimation
SiLK: Simple Learned Keypoints
Pierre Gleize, Weiyao Wang, Matt Feiszli
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| remove zero padding in convolutions (ablation) | SiLK · HPatches | homography estimation accuracy @1px: 0.59 -> 0.62 | Improved |
| remove zero padding in convolutions (ablation) | SiLK · ScanNet | rotation accuracy @5 deg: 93.7 -> 98.1 | Improved |
| remove zero padding in convolutions (ablation) | SiLK · ScanNet | translation accuracy @5cm: 75.2 -> 82.9 | Improved |
| replace VGGnp-4 by FPN backbone (ablation) | SiLK · HPatches | homography estimation accuracy @1px: 0.62 -> 0.55 | Hurt |
| replace VGGnp-4 by UNet backbone (ablation) | SiLK · HPatches | homography estimation accuracy @1px: 0.62 -> 0.41 | Hurt |
| shrink backbone to VGGnp-micro (76k params) (ablation) | SiLK · HPatches | homography estimation accuracy @1px: 0.62 -> 0.56; MMA 0.59 -> 0.40 | Mixed |
| shrink backbone to VGGnp-micro (ablation) | SiLK · HPatches | repeatability @1px: 0.62 -> 0.64 | No clear effect |
| reduce training descriptor map from 146x146 to 82x82 (ablation) | SiLK · HPatches | homography estimation accuracy @1px: 0.59 -> 0.58; training time 5h -> 1.7h | No clear effect |
| train on ScanNet instead of COCO (ablation) | SiLK · HPatches | mean matching accuracy @1px: 0.59 -> 0.48 | Hurt |
| train on ImageNet instead of COCO (ablation) | SiLK · ScanNet | translation accuracy @5cm: 82.9 -> 83.5 | No clear effect |
| increase top-k from 1k to 10k (ablation) | SiLK · HPatches | homography estimation accuracy @1px: 0.53 -> 0.62 | Improved |
| use SiLK with simple MNN and no context aggregation vs SuperGlue | SiLK · ScanNet relative pose | pose AUC @5 deg: 16.2 -> 18.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SuperGlue — SuperGlue: Learning Feature Matching With Graph Neural Networks (CVPR 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- X-Pose: Detecting Any Keypoints (ECCV 2024 · similarity 0.86)
- End-to-End Learning of Keypoint Detector and Descriptor for Pose Invariant 3D Matching (CVPR 2018 · similarity 0.86)
- Detect Any Keypoints: An Efficient Light-Weight Few-Shot Keypoint Detector (AAAI 2024 · similarity 0.85)
- Back to 3D: Few-Shot 3D Keypoint Detection with Back-Projected 2D Features (CVPR 2024 · similarity 0.84)
- Towards Complementary Keypoint Detection via Mixture of Detectors (NeurIPS 2026 · similarity 0.84)
- Key-Grid: Unsupervised 3D Keypoints Detection using Grid Heatmap Features (NeurIPS 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.
