Research map / Video / Action recognition
Two-Stage Active Learning for Efficient Temporal Action Segmentation
Yuhao Su, Ehsan Elhamifar
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 regularizers L_mu and L_sigma to contrastive prototype loss (ablation) | Action prototype learning in TAS · 50Salads | accuracy: 47.2 -> 57.8 | Improved |
| add regularizers L_mu and L_sigma to contrastive prototype loss (ablation) | Action prototype learning in TAS · GTEA | accuracy: 43.9 -> 47.6 | Improved |
| use L_cont alone instead of plain average-embedding prototypes (ablation) | Action prototype learning in TAS · 50Salads | accuracy: 50.0 -> 47.2 | Hurt |
| use VAS intra-video selection instead of split-random (ablation) | Two-stage AL with VPA · 50Salads | accuracy: 45.4 -> 57.8 | Improved |
| use VPA inter-video selection instead of random selection (ablation) | Two-stage AL with VAS · 50Salads | accuracy: 44.7 -> 57.8 | Improved |
| use proposed two-stage selection versus split-random baseline | TAS active learning · 50Salads (0.16% frames) | f1@50: 24.7 -> 32.9 | Improved |
| use proposed selection on CrossTask with 72% background frames | TAS active learning · CrossTask (0.16% frames) | f1@50: 11.2 -> 12.4 | Improved |
| add pseudo-labels for low-alignment-cost videos from 5th iteration | Two-stage AL on Breakfast · Breakfast (0.31% frames) | f1@50 as percent of full supervision: 90% -> close to 95% | Improved |
| keep adding pseudo-labels beyond 0.35% budget | Two-stage AL on Breakfast · Breakfast (>0.35% frames) | f1@50 | Hurt |
| use temporally separated frame labeling baselines instead of coreset | Intra-video selection baselines · all four datasets | TAS performance: 38.4 -> 49.0 acc (50Salads, coreset vs split-rand) | Improved |
| label 0.23% frames with our method versus all-video timestamp supervision (0.29%) | Active learning vs timestamp supervision · Breakfast | f1@50: 50.7 -> 54.3 (D-TSTAS vs ours) | Improved |
Closest papers by meaning
- Temporal Action Segmentation: An Analysis of Modern Techniques (TPAMI 2023 · similarity 0.92)
- FACT: Frame-Action Cross-Attention Temporal Modeling for Efficient Action Segmentation (CVPR 2024 · similarity 0.90)
- Unsupervised Action Segmentation via Fast Learning of Semantically Consistent Actoms (AAAI 2024 · similarity 0.89)
- Efficient and Effective Weakly-Supervised Action Segmentation via Action-Transition-Aware Boundary Alignment (CVPR 2024 · similarity 0.89)
- OnlineTAS: An Online Baseline for Temporal Action Segmentation (NeurIPS 2024 · similarity 0.89)
- Understanding Multi-Task Activities from Single-Task Videos (CVPR 2025 · 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.
