Research map / Label-efficient and robust learning / Noisy labels
Rethinking Pseudo-Label Guided Learning for Weakly Supervised Temporal Action Localization from the Perspective of Noise Correction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NoCo, a noise-correction framework for pseudo-label learning in weakly supervised temporal action localization, combining context-aware label augmentation with an online teacher-student framework (missing-instance compensation, ambiguous-instance correction, high-quality pseudo-label mining loss).
Problem
Pseudo-labels generated by weakly supervised TAL models are noisy (inaccurate boundaries, missed short actions, adjacent instances merged), and this noise degrades the fully supervised detection head trained on them.
Approach
Refine pseudo-label boundaries with context-aware weighted averaging of associated instances, then correct noise online in a teacher-student loop: the AIC module fixes over-complete/merged instances, the MIC module adds missing short instances from teacher predictions, and an HPM loss weights instances by quality. The noise correction is decoupled from the WTAL base model, so inference uses only t
Key results
- On THUMOS14, applying NoCo to ASM-Loc raises average mAP (0.3:0.7) from 35.8 to 42.1 in the ablation, and the ASM-Loc+NoCo variant reaches 42.1 vs 36.0 for CO2-Net-based gains of +6.0.
- On ActivityNet v1.2, ASM-Loc+NoCo achieves 30.7 average mAP vs 26.5 for ASM-Loc baseline.
- Inference speed rises from 1.64 to 6.67 video/s with FLOPS reduced from 124.79G to 20.12G.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply NoCo to ASM-Loc | ASM-Loc · ActivityNet v1.2 validation | AVG mAP (%): 26.5 -> 30.7 | Helped |
| Apply NoCo to CO₂-Net | CO₂-Net · ActivityNet v1.2 validation | AVG mAP (%): 26.4 -> 30.1 | Helped |
| Apply NoCo to LACP | LACP · ActivityNet v1.2 validation | AVG mAP (%): 26.8 -> 30.7 | Helped |
| Apply the full NoCo framework | WTAL methods on THUMOS14 · THUMOS14 test | average mAP | Helped |
| Add CALA to the pseudo-label retraining baseline (ablation) | ASM-Loc with TriDet retraining · THUMOS14 | AVG mAP (%): 35.8 -> 37.1 (+1.3) | Helped |
| Add teacher-student training to CALA (ablation) | ASM-Loc with TriDet retraining · THUMOS14 | AVG mAP (%): 37.1 -> 37.8 (+0.7) | Helped |
| Add AIC to CALA + teacher-student training (ablation) | ASM-Loc with TriDet retraining · THUMOS14 | AVG mAP (%): 37.8 -> 40.0 | Helped |
| Add MIC to CALA + teacher-student training (ablation) | ASM-Loc with TriDet retraining · THUMOS14 | AVG mAP (%): 37.8 -> 39.6 | Helped |
| Add both AIC and MIC to CALA + teacher-student training (ablation) | ASM-Loc with TriDet retraining · THUMOS14 | AVG mAP (%): 37.8 -> 41.9 | Helped |
| Add HPM to CALA + teacher-student training + AIC + MIC (ablation) | ASM-Loc with TriDet retraining · THUMOS14 | AVG mAP (%): 41.9 -> 42.1 (+0.2) | Helped |
| Add the full NoCo framework | ASM-Loc with TriDet retraining · THUMOS14 | speed (video/s): 1.64 -> 6.67 | Helped |
Surprising findings
- The teacher-student framework alone without correction modules gives only +0.7 mAP, while AIC and MIC contribute most of the gain.
Limitations
- Relies on a pre-trained WTAL base model to generate initial pseudo-labels, so noise originates from that model.
- Many hyperparameters (IoU thresholds, alpha, beta, confidence threshold) are introduced by the correction modules.
Experimental setup
Datasets: THUMOS14 ActivityNet v1.2
Benchmarks: THUMOS14 ActivityNet v1.2
Metrics: mAP@IoU average mAP inference speed (video/s) FLOPS
Baselines: ASM-Loc CO2-Net UM CoLA DELU CASE LACP BackTAL SSN SF-Net
Builds on: ASM-Loc TriDet DivideMix focal loss
Closest papers by meaning
- Improving Weakly Supervised Temporal Action Localization by Bridging Train-Test Gap in Pseudo Labels (CVPR 2023 · similarity 0.94)
- Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer (CVPR 2025 · similarity 0.93)
- Weakly-supervised Temporal Action Localization by Uncertainty Modeling (AAAI 2021 · similarity 0.89)
- Weakly-Supervised Temporal Action Localization by Inferring Salient Snippet-Feature (AAAI 2024 · similarity 0.88)
- HR-Pro: Point-Supervised Temporal Action Localization via Hierarchical Reliability Propagation (AAAI 2024 · similarity 0.88)
- Similar Modality Enhancement and Action Consistency Learning for Weakly Supervised Temporal Action Localization (AAAI 2025 · similarity 0.87)
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.