Research map / Detection and segmentation / Open-set and few-shot detection
From Keypoints to Object Landmarks via Self-Training Correspondence: A novel approach to Unsupervised Landmark Discovery
Dimitrios Mallis, Enrique Sánchez, Matthew Bell, Georgios Tzimiropoulos
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 |
|---|---|---|---|
| use SuperPoint instead of SIFT+ANMS for initial keypoints (ablation) | Stage 1 · CelebA | Forward-NME: 4.07 -> 3.25 | Improved |
| use SuperPoint instead of ORB+ANMS (ablation) | Stage 1 · BBCPose | Forward-NME: 17.31 -> 13.49 | Improved |
| use SuperPoint instead of ORB+ANMS (ablation) | Stage 1 · BBCPose | Backward-NME: 11.72 -> 13.55 | Hurt |
| over-segment features with M=100 clusters instead of M=30 (ablation) | Stage 1 · LS3D | Forward-NME: 10.26 -> 7.95 | Improved |
| increase clusters from M=100 to M=500 (ablation) | Stage 1 · LS3D | Forward-NME: 7.95 -> 9.53 | Hurt |
| increase clusters from M=100 to M=500 (ablation) | Stage 1 · LS3D | Backward-NME: 6.55 -> 6.19 | Improved |
| sample negatives from same image only instead of different clusters (ablation) | Clustering correspondence · LS3D | Forward-NME: 11.15 -> 7.95 | Improved |
| sample negatives from same image only instead of different clusters (ablation) | Equivariance correspondence · LS3D | Forward-NME: 9.56 -> 10.02 | Hurt |
| recover correspondence by clustering instead of equivariance (ablation) | Same-image negatives · LS3D | Forward-NME: 10.02 -> 7.95 | Improved |
| add Stage 2 landmark detector (ablation) | Stage 1 · LS3D | Forward-NME: 7.9 -> 5.2 | Improved |
| add Stage 2 landmark detector (ablation) | Stage 1 · CelebA | Forward-NME: 3.3 -> 3.2 | Improved |
| add flipping augmentation at training (ablation) | Stage 1 · CelebA | Forward-NME: 3.88 -> 3.32 | Improved |
Closest papers by meaning
- Unsupervised Learning of Object Landmarks via Self-Training Correspondence (NeurIPS 2020 · similarity 0.96)
- Object landmark discovery through unsupervised adaptation (NeurIPS 2019 · similarity 0.92)
- Unsupervised Learning of Object Landmarks by Factorized Spatial Embeddings (ICCV 2017 · similarity 0.90)
- Unsupervised Learning of Object Landmarks through Conditional Image Generation (NeurIPS 2018 · similarity 0.89)
- Pose-Guided Self-Training with Two-Stage Clustering for Unsupervised Landmark Discovery (CVPR 2024 · similarity 0.88)
- Unsupervised 3D Structure Inference from Category-Specific Image Collections (CVPR 2024 · 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.
