Research map / Video / Action recognition
Hierarchically Self-Supervised Transformer for Human Skeleton Representation Learning
Yuxiao Chen, Long Zhao, Jianbo Yuan, Yu Tian, Zhaoyang Xia, Shijie Geng, Ligong Han, Dimitris N. Metaxas
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 frame-level (F-TRS) pre-training (ablation) | Hi-TRS random init · NTU-60 xsub (supervised) | top-1 accuracy: 79.6 -> 80.8 | Improved |
| add clip-level (C-TRS) pre-training (ablation) | Hi-TRS random init · NTU-60 xsub (supervised) | top-1 accuracy: 79.6 -> 81.1 | Improved |
| add video-level (V-TRS) pre-training (ablation) | Hi-TRS random init · NTU-60 xsub (supervised) | top-1 accuracy / MPJPE: 79.6 -> 82.0 accuracy, 98.4 -> 88.1 MPJPE | Improved |
| add clip-level pre-training to frame-level (ablation) | Hi-TRS pre-trained at F · NTU-60 xsub (supervised) | top-1 accuracy: 80.8 -> 83.9 | Improved |
| add video-level pre-training to F+C (ablation) | Hi-TRS pre-trained at F+C · NTU-60 xsub (supervised) | top-1 accuracy: 83.9 -> 86.0 | Improved |
| pre-train on all three levels instead of F+V (ablation) | Hi-TRS · NTU-60 xsub (supervised) | MPJPE: 87.4 -> 85.6 | Improved |
| add video-level pre-training to F+C for detection (ablation) | Hi-TRS · PKUMMD Part I action detection | mAP_a / mAP_v: mAP_a 55.6 -> 58.4, mAP_v 65.1 -> 66.3 | Improved |
| add F+C pre-training for detection (ablation) | Hi-TRS random init · PKUMMD Part I action detection | mAP_a / mAP_v: mAP_a 53.4 -> 55.6, mAP_v 63.2 -> 65.1 | Improved |
| pre-train on NTU-60 then fine-tune on PKU Part II | Hi-TRS random init · PKUMMD Part II (transfer) | MPJPE / accuracy: MPJPE 105.4 -> 94.2, accuracy 50.9 -> 55.0 | Improved |
| pre-train on larger NTU-120 instead of NTU-60 | Hi-TRS · PKUMMD Part I (transfer) | mAP_a / mAP_v: mAP_a 55.2 -> 57.9, mAP_v 66.6 -> 67.3 | Improved |
| fuse three streams (joint, motion, bone) instead of single stream | Hi-TRS · NTU-60 xsub (supervised) | top-1 accuracy: 86.0 -> 90.0 | Improved |
| use only 1% labeled data vs strong baselines | Hi-TRS · NTU-60 xsub semi-supervised (1%) | top-1 accuracy: Hi-TRS 39.1 vs SCC-3S 48.3, CrosSCLR-3S 51.1 | Hurt |
Closest papers by meaning
- Self-Supervised Skeleton-Based Action Representation Learning: A Benchmark and Beyond (IJCV 2026 · similarity 0.91)
- Hierarchical Contrast for Unsupervised Skeleton-Based Action Representation Learning (AAAI 2023 · similarity 0.91)
- SkeletonMAE: Graph-based Masked Autoencoder for Skeleton Sequence Pre-training (ICCV 2023 · similarity 0.91)
- Self-Supervised 3D Skeleton Action Representation Learning With Motion Consistency and Continuity (ICCV 2021 · similarity 0.89)
- Heterogeneous Skeleton-Based Action Representation Learning (CVPR 2025 · similarity 0.87)
- Stitch, Contrast, and Segment: Learning a Human Action Segmentation Model Using Trimmed Skeleton Videos (AAAI 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.
