Research map / Video / Action recognition
Towards Efficient General Feature Prediction in Masked Skeleton Modeling
Shengkai Sun, Zefan Zhang, Jianfeng Dong, Zhiyong Cheng, Xiaojun Chang, Meng Wang
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 |
|---|---|---|---|
| add motion-difference input for TGN instead of raw joints (ablation) | GFP · NTU-60 x-view | top-1 accuracy: 90.9 -> 92.0 | Improved |
| use unmasked motion input instead of masked joint input for TGN (ablation) | GFP · NTU-60 x-view | top-1 accuracy: 90.3 -> 92.0 | Improved |
| use 3-layer TGN instead of shallow 2-layer (ablation) | GFP · NTU-60 x-view | top-1 accuracy: 92.0 -> 91.8 | No clear effect |
| use 3-layer TGN instead of wider 4-layer (ablation) | GFP · NTU-60 x-view | top-1 accuracy: 92.0 -> 91.3 | No clear effect |
| use 3-layer projector instead of 2 layers (ablation) | GFP · NTU-60 x-view | top-1 accuracy: 91.3 -> 92.0 | Improved |
| use 3-layer projector instead of 4 layers (ablation) | GFP · NTU-60 x-view | top-1 accuracy: 91.6 -> 92.0 | Improved |
| add global and local hierarchical targets via TGN rather than removing them (ablation) | GFP · NTU-60 x-sub | top-1 accuracy | Improved |
| add TGN-learned targets instead of direct low-level target prediction (ablation) | GFP · NTU-60 x-sub | top-1 accuracy | Improved |
| replace SkeletonMAE-style reconstruction with hierarchical feature prediction | GFP vs SkeletonMAE · NTU-60 pretraining | training speedup: 1x -> 6.2x | Improved |
| replace patch-level EMA targets (S-JEPA) with lightweight TGN | GFP vs S-JEPA · NTU-60 x-view | top-1 accuracy: 89.8 -> 92.0 | Improved |
| replace MAMP motion targets with GFP high-level targets | GFP vs MAMP · NTU-60 retrieval x-view | top-1 accuracy: 70.0 -> 87.1 | Improved |
| use GFP instead of S-JEPA | GFP vs S-JEPA · NTU-120 x-sub | top-1 accuracy: 79.6 -> 79.1 | Hurt |
Closest papers by meaning
- Exploring Adaptive Masked Reconstruction for Self-Supervised Skeleton-Based Action Recognition (CVPR 2026 · similarity 0.91)
- Less is More: Compact-Token Masked Feature Learning for Skeleton Representation Learning (NeurIPS 2026 · similarity 0.90)
- SkeletonMAE: Graph-based Masked Autoencoder for Skeleton Sequence Pre-training (ICCV 2023 · similarity 0.90)
- MaskCLR: Attention-Guided Contrastive Learning for Robust Action Representation Learning (CVPR 2024 · similarity 0.88)
- Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.87)
- Heterogeneous Skeleton-Based Action Representation Learning (CVPR 2025 · similarity 0.87)
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.
