Research map / Label-efficient and robust learning / Self-supervised and contrastive
S-JEPA: A Joint Embedding Predictive Architecture for Skeletal Action Recognition
Mohamed Abdelfattah, Alexandre Alahi
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 |
|---|---|---|---|
| update target encoder with EMA of view encoder (ablation) | S-JEPA · NTU60 XSub, linear eval | accuracy (%): 1.6 -> 85.3 | Improved |
| add geometric transformations to create diverse views (ablation) | S-JEPA · NTU60 XSub, linear eval | accuracy (%): 84.6 -> 85.3 | Improved |
| use cross-entropy loss over softmax distributions instead of MSE (ablation) | S-JEPA · NTU60 XSub, linear eval | accuracy (%): 83.7 -> 85.3 | Improved |
| add centering of target representations (ablation) | S-JEPA · NTU60 XSub, linear eval | accuracy (%): 84.5 -> 85.3 | Improved |
| mask output of target encoder instead of its input (ablation) | S-JEPA · NTU60 XSub, linear eval | accuracy (%): 78.1 -> 85.3 | Improved |
| pre-train with S-JEPA before fine-tuning vs training from scratch | vanilla Transformer · NTU60 XSub fine-tuning | accuracy (%): +10% | Improved |
| pre-train with S-JEPA vs from scratch with 1% labels | vanilla Transformer · NTU60 XSub semi-supervised | accuracy (%): +28.7 points | Improved |
| predict latent representations instead of masked motion (MAMP) | masked skeleton pretraining · NTU60 XSub fine-tuning | accuracy (%): 93.1 -> 93.1 | No clear effect |
| predict latent representations instead of masked motion (MAMP) | masked skeleton pretraining · NTU60 XView 10% labels | accuracy (%): 91.5 -> 91.4 | Hurt |
| predict latent representations instead of masked motion (MAMP) | masked skeleton pretraining · transfer to PKU-II | accuracy (%): 70.6 -> 71.4 (NTU60 source) | Improved |
| set EMA update rate to 0.9999 (ablation) | S-JEPA · NTU60 XSub, linear eval | accuracy (%): 83.6 (0.999) -> 85.3 (0.9999) | Improved |
Closest papers by meaning
- Heterogeneous Skeleton-Based Action Representation Learning (CVPR 2025 · similarity 0.88)
- MaskCLR: Attention-Guided Contrastive Learning for Robust Action Representation Learning (CVPR 2024 · similarity 0.88)
- Neuron: Learning Context-Aware Evolving Representations for Zero-Shot Skeleton Action Recognition (CVPR 2025 · similarity 0.87)
- Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.86)
- Skeleton-based Action Recognition with Non-linear Dependency Modeling and Hilbert-Schmidt Independence Criterion (AAAI 2025 · similarity 0.86)
- SAMoR: Motion Modelling for Articulated Objects of Any Skeleton and Topology (NeurIPS 2026 · similarity 0.86)
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.
