Research map / Video / Action recognition
What do I Annotate Next? An Empirical Study of Active Learning for Action Localization
Fabian Caba Heilbron, Joon-Young Lee, Hailin Jin, Bernard Ghanem
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 |
|---|---|---|---|
| replace random sampling with LAL selection | Active learner · ActivityNet-Selection active learning (150 classes) | labeled data to reach 80% of final mAP: 65% -> 35% | Improved |
| replace random sampling with MCLE selection | Active learner · ActivityNet-Selection active learning (150 classes) | labeled data to reach 80% of final mAP: 65% -> 38% | Improved |
| replace uncertainty sampling with LAL selection | Active learner · ActivityNet-Selection active learning (150 classes) | labeled data to reach 80% of final mAP: 42% -> 35% | Improved |
| replace uncertainty sampling with MCLE selection | Active learner · ActivityNet-Selection active learning (150 classes) | labeled data to reach 80% of final mAP: 42% -> 38% | Improved |
| increase sampling batch size from 1 to 64 (ablation) | MCLE · ActivityNet-Selection active learning (150 classes) | AULC: 0.75% -> 0.65% | Hurt |
| increase sampling batch size to 64 (ablation) | LAL · ActivityNet-Selection active learning (150 classes) | AULC: drops only 0.05 | No clear effect |
| increase sampling batch size to 32 or more (ablation) | Uncertainty sampling · ActivityNet-Selection active learning (150 classes) | AULC | Hurt |
| inject 5% noise into oracle annotations (ablation) | LAL · ActivityNet-Selection active learning (150 classes) | AULC | No clear effect |
| train on Kinetics-Localization instead of original Kinetics | Stage-Wise localization model · Kinetics-Localization validation set | average mAP (%): 8.2 -> 21.3 | Improved |
| train on Kinetics-Localization instead of original Kinetics | End-to-End SSN · Kinetics-Localization validation set | average mAP (%): 8.3 -> 22.8 | Improved |
| raise tIoU threshold from 0.1 to 0.2 | End-to-End SSN trained on Kinetics · Kinetics-Localization validation set | mAP (%): 59.4 -> 40.1 | Hurt |
Closest papers by meaning
- Temporal Action Localization in the Deep Learning Era: A Survey (TPAMI 2023 · similarity 0.90)
- Learning to Localize Actions from Moments (ECCV 2020 · similarity 0.90)
- Two-Stage Active Learning for Efficient Temporal Action Segmentation (ECCV 2024 · similarity 0.89)
- HR-Pro: Point-Supervised Temporal Action Localization via Hierarchical Reliability Propagation (AAAI 2024 · similarity 0.89)
- Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer (CVPR 2025 · similarity 0.88)
- Energy vs. Noise: Towards Robust Temporal Action Localization in Open-World (AAAI 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.
