Research map / Video / Action recognition
Learning Human Action Recognition Representations Without Real Humans
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| remove supervised alignment stage (MAE only) (ablation) | MAE-Align pretrained on Kinetics · 6 downstream tasks, finetune | average top-1 accuracy (%): 68.1 -> 59.9 | Hurt |
| remove MAE stage (alignment only) (ablation) | MAE-Align pretrained on Kinetics · 6 downstream tasks, finetune | average top-1 accuracy (%): 68.1 -> 32.3 | Hurt |
| remove MAE stage (alignment only) (ablation) | MAE-Align pretrained on Synthetic · 6 downstream tasks, finetune | average top-1 accuracy (%): 64.5 -> 31.1 | Hurt |
| remove alignment stage (MAE only) (ablation) | MAE-Align on Synthetic · 6 downstream tasks, linear probe | average top-1 accuracy (%): 42.9 -> 27.8 | Hurt |
| add alignment after MAE on NH Kinetics (ablation) | MAE on NH Kinetics · 6 downstream tasks, linear probe | average top-1 accuracy (%): 28.4 -> 42.4 | Helped |
| replace real Kinetics with synthetic data for MAE pretraining (ablation) | MAE-Align · 6 downstream tasks, finetune | average top-1 accuracy (%): 68.1 -> 64.5 | Hurt |
| MAE-Align instead of TSN (RN50) on synthetic data (ablation) | synthetic pretraining · 6 downstream tasks, finetune | average top-1 accuracy (%): 54.9 -> 64.5 | Helped |
| MAE-Align instead of TSN (RN50) on synthetic data (ablation) | synthetic pretraining · 6 downstream tasks, linear probe | average top-1 accuracy (%): 19.1 -> 42.9 | Helped |
| add No-Human Kinetics to synthetic data (PPMA) (ablation) | MAE-Align w/ Synthetic · 6 downstream tasks, finetune | average top-1 accuracy (%): 64.5 -> 67.0 | Helped |
| add No-Human Kinetics to synthetic data (PPMA) (ablation) | MAE-Align w/ Synthetic · 6 downstream tasks, linear probe | average top-1 accuracy (%): 42.9 -> 47.9 | Helped |
| align on NH Kinetics plus Synthetic instead of NH Kinetics alone (ablation) | PPMA · 6 downstream tasks, finetune | average top-1 accuracy (%): 65.6 -> 67.0 | Helped |
| combine models by averaging weights of two separately aligned models instead of aligning on mixed data (ablation) | PPMA · 6 downstream tasks, finetune | average top-1 accuracy (%): 67.0 -> 66.5 | Hurt |
Closest papers by meaning
- Procedural Generation of Videos to Train Deep Action Recognition Networks (CVPR 2017 · similarity 0.88)
- Large-Scale Weakly-Supervised Pre-Training for Video Action Recognition (CVPR 2019 · similarity 0.87)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.85)
- Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training (NeurIPS 2024 · similarity 0.85)
- MVHumanNet: A Large-scale Dataset of Multi-view Daily Dressing Human Captures (CVPR 2024 · similarity 0.85)
- HumanVid: Demystifying Training Data for Camera-controllable Human Image Animation (NeurIPS 2024 · similarity 0.85)
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.