Research map / 3D vision / Human pose estimation
Slim DensePose: Thrifty Learning From Sparse Annotations and Motion Cues
Natalia Neverova, James Thewlis, Riza Alp Guler, Iasonas Kokkinos, Andrea Vedaldi
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 sparse UV annotations on more images rather than complete annotations on a small image subset (ablation) | Hourglass DensePose model · DensePose-COCO | 5 cm: 36.27 -> 47.68 | Improved |
| Add part labels across images while sparsifying UV labels (ablation) | Hourglass DensePose model · DensePose-COCO | 5 cm: 37.65 -> 40.80 | Improved |
| Add keypoint annotations to 1% UV supervision (ablation) | Hourglass DensePose model · DensePose-COCO | 5 cm: 37.65 -> 39.17 | Improved |
| Use segmentation-only supervision instead of UV correspondence labels (ablation) | Hourglass DensePose model · DensePose-COCO | 5 cm: 55.78 -> 3.53 | Hurt |
| Pretrain on COCO before fine-tuning on DensePose-Track | Hourglass DensePose model · DensePose-Track | 5 cm: 21.06 -> 45.57 | Improved |
| Fine-tune on the union of COCO and Track after COCO pretraining | Hourglass DensePose model · DensePose-Track | 5 cm: 45.57 -> 46.04 | Improved |
| Add ground-truth propagation using optical flow (ablation) | Hourglass DensePose model · DensePose-Track | 20 cm: 59.54 -> 75.95 | Improved |
| Add equivariance using optical flow (ablation) | Hourglass DensePose model · DensePose-Track | 20 cm: 59.54 -> 62.22 | Improved |
| Combine optical-flow GT propagation and equivariance (ablation) | Hourglass DensePose model · DensePose-Track | 20 cm: 59.54 -> 77.20 | Improved |
| Use real optical flow instead of synthetic TPS motion (ablation) | Hourglass DensePose model with GT propagation and equivariance · DensePose-Track | 20 cm: 62.71 -> 77.20 | Improved |
| Add optical-flow GT propagation and equivariance after COCO pretraining (ablation) | Hourglass DensePose model · DensePose-Track | 20 cm: 83.71 -> 87.56 | Improved |
| Train jointly on COCO and Track without staged fine-tuning (ablation) | Hourglass DensePose model · DensePose-Track | 5 cm: 44.89 -> 41.76 | Hurt |
Closest papers by meaning
- DensePose: Dense Human Pose Estimation in the Wild (CVPR 2018 · similarity 0.88)
- Learning Temporal Pose Estimation from Sparsely-Labeled Videos (NeurIPS 2019 · similarity 0.88)
- ProxyPose: 6-DoF Pose Tracking via Video-to-Video Translation (NeurIPS 2026 · similarity 0.85)
- Dynamic Camera Poses and Where to Find Them (CVPR 2025 · similarity 0.85)
- Can Generative Video Models Help Pose Estimation? (CVPR 2025 · similarity 0.85)
- Fast Encoder-Based 3D from Casual Videos via Point Track Processing (NeurIPS 2024 · similarity 0.85)
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.
