Research map / 3D vision / Human pose estimation
3D Human Pose Estimation in Video With Temporal Convolutions and Semi-Supervised Training
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 |
|---|---|---|---|
| add temporal convolutions over 243 frames instead of using a single-frame model (ablation) | 3D pose estimation model · Human3.6M | MPJPE (Protocol 1): 51.8 -> 46.8 | Helped |
| use causal rather than full temporal convolutions (ablation) | 3D pose estimation model · Human3.6M | MPJPE (Protocol 1): 46.8 -> 49.0 | Hurt |
| increase receptive field from 27 to 81 frames (ablation) | temporal convolutional model · Human3.6M, ground-truth 2D poses | MPJPE: 40.6 -> 38.7 | Helped |
| increase receptive field from 81 to 243 frames (ablation) | temporal convolutional model · Human3.6M, ground-truth 2D poses | MPJPE: 38.7 -> 37.8 | Helped |
| replace dilated convolutions with regular dense convolutions (ablation) | 27-frame temporal convolutional model · Human3.6M, fine-tuned CPN detections | MPJPE: 48.8 -> 50.4 (+1.6 mm) | Hurt |
| remove test-time horizontal flip augmentation (ablation) | top-performing temporal convolutional model · Human3.6M | MPJPE: 46.8 -> 47.7 | Hurt |
| also remove train-time augmentation (ablation) | temporal convolutional model without test-time augmentation · Human3.6M | MPJPE: +1.5 mm | Hurt |
| add the bone-length L2 constraint to semi-supervised back-projection (ablation) | semi-supervised 3D pose estimator · Human3.6M, 1% of S1 labeled | MPJPE: 91.3 mm -> 78.1 mm | Helped |
| add semi-supervised back-projection training | 3D pose estimator · Human3.6M, fewer than 5K labeled frames | N-MPJPE: 9-10.4 mm | Helped |
| add semi-supervised back-projection training | 3D pose estimator · Human3.6M, full frame rate | MPJPE: up to 14.7 mm | Helped |
| use ground-truth 2D poses instead of CPN detections in semi-supervised training | 3D pose estimator · Human3.6M, 1% of S1 | MPJPE: up to 22.6 mm | Helped |
Closest papers by meaning
- Self-Supervised 3D Human Pose Estimation via Part Guided Novel Image Synthesis (CVPR 2020 · similarity 0.92)
- 3D Human Pose Estimation Using Spatio-Temporal Networks with Explicit Occlusion Training (AAAI 2020 · similarity 0.90)
- Weakly-Supervised 3D Human Pose Learning via Multi-View Images in the Wild (CVPR 2020 · similarity 0.90)
- Neural View Synthesis and Matching for Semi-Supervised Few-Shot Learning of 3D Pose (NeurIPS 2021 · similarity 0.90)
- Multiview-Consistent Semi-Supervised Learning for 3D Human Pose Estimation (CVPR 2020 · similarity 0.90)
- 3D Human Pose Estimation From a Single Image via Distance Matrix Regression (CVPR 2017 · similarity 0.90)
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.