Research map / 3D vision / Human pose estimation
PoseFormerV2: Exploring Frequency Domain for Efficient and Robust 3D Human Pose Estimation
Read the paperCheck 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 |
|---|---|---|---|
| sample central frames from a longer sequence without adding frequency features (ablation) | PoseFormerV1 · Human3.6M Protocol 1 | MPJPE: 49.9 -> 49.9 | No clear effect |
| append low-frequency DCT coefficient embeddings from the full sequence (ablation) | PoseFormerV1 · Human3.6M Protocol 1 | MPJPE: 49.9 -> 47.1 (2.8↓) | Helped |
| replace the time-feature vanilla MLP with FreqMLP (ablation) | PoseFormerV1 with DCT features · Human3.6M Protocol 1 | MPJPE: 47.1 -> 46.0 (3.9↓ from original) | Helped |
| increase retained DCT coefficients from 1 to 3 (ablation) | PoseFormerV2 with f=1 · Human3.6M Protocol 1, full length 27 | MPJPE: 51.1 -> 48.7 (2.4↓) | Helped |
| increase central input frames from 1 to 3 while keeping one DCT coefficient (ablation) | PoseFormerV2 · Human3.6M Protocol 1, full length 27 | MPJPE: 51.1 -> 50.1 (1.0↓) | Helped |
| increase frames and retained DCT coefficients to 3 each (ablation) | PoseFormerV2 · Human3.6M Protocol 1, full length 27 | MPJPE: 51.1 -> 47.9 (3.2↓) | Helped |
| use 81-frame context with 3 central frames and 3 DCT coefficients instead of the frequency-only baseline (ablation) | PoseFormerV2 · Human3.6M Protocol 1 | MPJPE: 49.7 -> 47.1 (2.6↓) | Helped |
| add low-frequency DCT sequence representation | PoseFormerV2 · Human3.6M with noisy 2D detections | performance drop as Gaussian noise increases | Helped |
| use low-frequency DCT coefficients to extend receptive field | PoseFormerV2 · Human3.6M Protocol 1 | MPJPE: 47.8 -> 46.0 | Helped |
| add low-frequency DCT temporal features to MixSTE | MixSTE · Human3.6M Protocol 1 | MPJPE: 46.2 -> 45.3 | Helped |
| add frequency representation and central-frame features to a frequency-only baseline (ablation) | PoseFormerV2 · Human3.6M Protocol 1, full length 81, f=3, n=3 | MPJPE: 49.7 -> 47.1 (2.6↓) | Helped |
| use minor Gaussian input noise | PoseFormerV2 · Human3.6M | MPJPE: sigma = 3 | Helped |
Closest papers by meaning
- Optimizing Human Pose Estimation Through Focused Human and Joint Regions (AAAI 2025 · similarity 0.90)
- SVTformer: Spatial-View-Temporal Transformer for Multi-View 3D Human Pose Estimation (AAAI 2025 · similarity 0.90)
- Lifting by Image – Leveraging Image Cues for Accurate 3D Human Pose Estimation (AAAI 2024 · similarity 0.89)
- HiPART: Hierarchical Pose AutoRegressive Transformer for Occluded 3D Human Pose Estimation (CVPR 2025 · similarity 0.89)
- FusionFormer: A Concise Unified Feature Fusion Transformer for 3D Pose Estimation (AAAI 2024 · similarity 0.88)
- ExtPose: Robust and Coherent Pose Estimation by Extending ViTs (ICML 2025 · similarity 0.88)
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.