Research map / Video / Action recognition
HR-Pro: Point-Supervised Temporal Action Localization via Hierarchical Reliability Propagation
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 HR-Pro, a point-supervised temporal action localization framework that propagates reliable point-annotation cues at snippet level (memory-based contrastive learning with a reliability-aware attention block) and instance level (point-based reliable proposals for completeness scoring and boundary regression).
Problem
Point-supervised temporal action localization (PSTAL) under-uses the reliability of point annotations, since existing methods optimize either at snippet level or instance level only, yielding noisy snippet scores and unreliable proposal scores.
Approach
An online-updated per-class prototype memory of reliable snippets feeds a Reliability-aware Attention Block that attends over intra- and inter-video snippets, plus a contrastive loss; at instance level, reliable proposals (highest OIC score containing each point) guide completeness score and boundary regression heads.
Key results
- HR-Pro reaches 60.3% average mAP over IoU 0.1:0.7 on THUMOS14, up 6.5 points over the prior best point-supervised method (CRRC-Net/LACP-level baselines, 53.8%).
- On THUMOS14 average mAP over 0.3:0.7, HR-Pro scores 51.1% vs. 52.0% for fully-supervised AFSD.
- Ablation: baseline 49.4% rises to 60.3% with contrastive loss, RAB, and instance-level losses combined.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add L_contra to the baseline objective (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 49.4 -> 51.5 | Helped |
| Add RAB to the model with L_base and L_contra (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 51.5 -> 54.7 | Helped |
| Add L_reg and L_score to the snippet-level method (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 54.7 -> 57.8 | Helped |
| Add reliable proposals to the method with L_score (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 57.1 -> 58.1 | Helped |
| Add negative proposals to the method with L_reg, L_score, and reliable proposals (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 59.1 -> 60.3 | Helped |
| Add L_contra to the baseline and instance-level components without RAB (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 53.9 -> 56.2 | Helped |
| Add RAB to the method with L_contra and instance-level components (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 56.2 -> 60.3 | Helped |
| Use uniform rather than Gaussian point annotation distribution (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 60.3 -> 58.8 | Hurt |
| Use center rather than Gaussian point annotation distribution (ablation) | HR-Pro · THUMOS14 | mAP@IoU [0.1:0.7]: 60.3 -> 60.8 | No clear effect |
Surprising findings
- Uniform point distribution yields slightly lower performance (58.8%) than center (60.8%) or Gaussian (60.3%) point sampling, which the authors attribute to less discriminative points being annotated.
Limitations
- Performance at high IoU thresholds (0.6-0.7) lags some methods, e.g., Zhou et al. at 0.6 and DELU at 0.7 on THUMOS14.
- Gains on ActivityNet1.3 are modest relative to fully-supervised methods (27.1% vs. 25.1% for LACP on the 0.5:0.95 average).
Experimental setup
Datasets: THUMOS14 GTEA BEOID ActivityNet 1.3 Kinetics-400
Benchmarks: THUMOS14 GTEA BEOID ActivityNet 1.3
Metrics: mAP@IoU average mAP over IoU thresholds
Baselines: SF-Net Ju et al. LACP CRRC-Net CoLA Zhou et al. AFSD DELU
Builds on: LACP BaS-Net Ju et al. point-supervised TAL OIC score soft-NMS
Closest papers by meaning
- Weakly Supervised Temporal Action Localization via Representative Snippet Knowledge Propagation (CVPR 2022 · similarity 0.91)
- Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer (CVPR 2025 · similarity 0.90)
- Realigning Confidence with Temporal Saliency Information for Point-Level Weakly-Supervised Temporal Action Localization (CVPR 2024 · similarity 0.90)
- Improving Weakly Supervised Temporal Action Localization by Bridging Train-Test Gap in Pseudo Labels (CVPR 2023 · similarity 0.90)
- Weakly-Supervised Temporal Action Localization by Inferring Salient Snippet-Feature (AAAI 2024 · similarity 0.90)
- Rethinking Pseudo-Label Guided Learning for Weakly Supervised Temporal Action Localization from the Perspective of Noise Correction (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.