Research map / Label-efficient and robust learning / Self-supervised and contrastive
HaLP: Hallucinating Latent Positives for Skeleton-Based Self-Supervised Learning of Actions
Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, Rama Chellappa
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 latent hallucinated positives to single-modality contrastive training | MoCo skeleton baseline · NTU-60 x-sub linear evaluation | accuracy: +2.3% | Improved |
| Add latent hallucinated positives to CMD multi-modal training | CMD · NTU-60 x-sub linear evaluation | accuracy: 79.8 -> 82.1 | Improved |
| Add latent hallucinated positives to single-modality contrastive training | MoCo skeleton baseline · NTU-120 x-sub linear evaluation | accuracy: +2.3% | Improved |
| Add latent hallucinated positives to single-modality contrastive training | MoCo skeleton baseline · PKU-II x-sub linear evaluation | accuracy: +4.5% | Improved |
| Use query-key geodesic interpolation to generate positives (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 77.92% | Hurt |
| Use only the hardest positives at t_c = t* instead of a range of hardness (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 79.4% | Hurt |
| Use a query feature as the hallucination anchor instead of a key feature (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 77.9% | Hurt |
| Use 20 prototypes instead of 10, 40, or 60 (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 79.5 -> 79.7 | Improved |
| Replace the rank filter with variant-1 filtering (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 79.7 -> 79.5 | Hurt |
| Replace the rank filter with variant-2 filtering (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 79.7 -> 79.6 | Hurt |
| Increase the HaLP loss weight from μ=0.5 to μ=1.0 (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 78.8 -> 79.7 | Improved |
| Generate 100 positives rather than 0 (ablation) | HaLP · NTU-60 x-sub linear evaluation | accuracy: 77.96 -> 79.71 | Improved |
Closest papers by meaning
- Hierarchical Consistent Contrastive Learning for Skeleton-Based Action Recognition with Growing Augmentations (AAAI 2023 · similarity 0.92)
- MaskCLR: Attention-Guided Contrastive Learning for Robust Action Representation Learning (CVPR 2024 · similarity 0.90)
- Stitch, Contrast, and Segment: Learning a Human Action Segmentation Model Using Trimmed Skeleton Videos (AAAI 2025 · similarity 0.89)
- Neuron: Learning Context-Aware Evolving Representations for Zero-Shot Skeleton Action Recognition (CVPR 2025 · similarity 0.88)
- Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.88)
- Recovering Complete Actions for Cross-dataset Skeleton Action Recognition (NeurIPS 2024 · 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.
