Research map / Label-efficient and robust learning / Self-supervised and contrastive
Hierarchical Consistent Contrastive Learning for Skeleton-Based Action Recognition with Growing Augmentations
Jiahang Zhang, Lilang Lin, Jiaying Liu
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 |
|---|---|---|---|
| Use HiCLR's growing augmentation and asymmetric hierarchical consistency with strong augmentations | ST-GCN · NTU60, linear evaluation | accuracy: 78.9 -> 80.4; 83.8 -> 85.5 | Improved |
| Use HiCLR's growing augmentation and asymmetric hierarchical consistency with strong augmentations | DSTA-Net · NTU120, linear evaluation | accuracy: 66.0 -> 67.3; 68.7 -> 69.9 | Improved |
| Add Random Mask in HiCLR's strong-augmentation branch (ablation) | Joint stream · NTU60 xsub, linear evaluation | accuracy: 68.3 -> 77.6 | Improved |
| Add Random Mask in HiCLR's strong-augmentation branch (ablation) | Joint stream · NTU60 xview, linear evaluation | accuracy: 76.4 -> 82.0 | Improved |
| Add Drop/Add Edges as the strong augmentation (ablation) | Bone stream · NTU60 xsub, linear evaluation | accuracy: 69.4 -> 76.9 | Improved |
| Use Random Mask before weaker augmentations rather than after them (ablation) | HiCLR, joint stream · NTU60 xsub, linear evaluation | accuracy: 77.6 -> 74.7 | Hurt |
| Combine BA and NA into one branch and use two branches (ablation) | HiCLR, joint stream · NTU60 xsub, linear evaluation | accuracy: 77.6 -> 74.0 | Hurt |
| Combine NA and Mask into one branch and use two branches (ablation) | HiCLR, joint stream · NTU60 xsub, linear evaluation | accuracy: 77.6 -> 76.7 | Hurt |
| Use KL divergence rather than cosine similarity for hierarchical consistency (ablation) | HiCLR, joint stream · NTU60 xsub, linear evaluation | accuracy: 75.8% -> 77.6% | Improved |
| Use KL divergence rather than L1 for hierarchical consistency (ablation) | HiCLR, joint stream · NTU60 xsub, linear evaluation | accuracy: 73.6% -> 77.6% | Improved |
| Apply Random Mask directly, without the hierarchical design (ablation) | Baseline, joint stream · NTU60 xsub, linear evaluation | accuracy: 68.3% -> 56.7% | Hurt |
| Use the hierarchical design with Random Mask instead of applying it directly (ablation) | Baseline, joint stream · NTU60 xsub, linear evaluation | accuracy: 56.7% -> 77.6% | Improved |
Closest papers by meaning
- Hierarchical Contrast for Unsupervised Skeleton-Based Action Representation Learning (AAAI 2023 · similarity 0.93)
- HaLP: Hallucinating Latent Positives for Skeleton-Based Self-Supervised Learning of Actions (CVPR 2023 · similarity 0.92)
- Beyond Binary Contrast: Modeling Continuous Skeleton Action Spaces with Transitional Anchors (CVPR 2026 · similarity 0.91)
- MaskCLR: Attention-Guided Contrastive Learning for Robust Action Representation Learning (CVPR 2024 · similarity 0.90)
- Skeleton-based Action Recognition with Non-linear Dependency Modeling and Hilbert-Schmidt Independence Criterion (AAAI 2025 · similarity 0.88)
- 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.
