Research map / 3D vision / Human pose estimation
HumanGPS: Geodesic PreServing Feature for Dense Human Correspondences
Check 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 binary triplet supervision with geodesic-structured GPS losses (ablation) | Dense human correspondence features · SMPL intra-subject, non-occluded | EPE: 9.14 -> 7.12 | Helped |
| Add dense geodesic loss to consistency and sparse ordinal losses (ablation) | HumanGPS · SMPL intra-subject, non-occluded | EPE: 8.17 -> 7.50 | Helped |
| Add cross-view dense geodesic loss to consistency, sparse ordinal, and dense geodesic losses (ablation) | HumanGPS · SMPL intra-subject, non-occluded | EPE: 7.50 -> 7.32 | Helped |
| Apply multi-scale supervision (ablation) | HumanGPS · SMPL intra-subject, non-occluded | EPE: 7.32 -> 7.12 | Helped |
| Use consistency and dense geodesic loss without sparse ordinal or cross-view dense terms (ablation) | HumanGPS · SMPL intra-subject, non-occluded | EPE: 7.67 vs 7.12 | Hurt |
| Add cross-view sparse ordinal geodesic loss (ablation) | HumanGPS · Cross-view correspondence training | correspondence accuracy | No clear effect |
| Increase feature dimension from 16 to 32 or 64 (ablation) | HumanGPS · SMPL intra-subject, non-occluded | EPE: 7.12 -> 6.87; 7.12 -> 6.78 | Helped |
| Reduce feature dimension from 16 to 8 (ablation) | HumanGPS · SMPL intra-subject, non-occluded | EPE: 7.12 -> 9.34 | Hurt |
| Add HumanGPS features to PWC-Net | PWC-Net · SMPL intra-subject, all pixels | EPE: 12.82 -> 10.86 | Helped |
| Add HumanGPS features to RAFT | RAFT · SMPL inter-subject, all pixels | EPE: 66.04 -> 17.61 | Helped |
| Use HumanGPS features rather than the augmented feature extractor alone | PWC-Net · Single-Human Optical Flow (SHOF) | AEPE: 0.1411 -> 0.1239 | Helped |
| Use HumanGPS-pretrained features for DensePose regression | ResNet-101 FCN* · DensePose COCO, 5 cm | accuracy: 51.32 -> 53.01 | Helped |
Closest papers by meaning
- DensePose: Dense Human Pose Estimation in the Wild (CVPR 2018 · similarity 0.89)
- Dense Human Body Correspondences Using Convolutional Networks (CVPR 2016 · similarity 0.89)
- BodyMap: Learning Full-Body Dense Correspondence Map (CVPR 2022 · similarity 0.87)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.85)
- MeshPose: Unifying DensePose and 3D Body Mesh Reconstruction (CVPR 2024 · similarity 0.84)
- Learning Affine Correspondences by Integrating Geometric Constraints (CVPR 2025 · 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.