Research map / Video / Action recognition
Online Temporal Action Localization with Memory-Augmented Transformer
Youngkil Song, Dongkeun Kim, Minsu Cho, Suha Kwak
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 |
|---|---|---|---|
| use two separate decoders for start and end prediction instead of one (ablation) | MATR · THUMOS14 | average mAP: 42.7 -> 49.5 | Improved |
| use MATR flag-token memory instead of MAT or E2E-LOAD memory modules (ablation) | MATR · THUMOS14 | average mAP: 46.9 / 47.9 -> 49.5 | Improved |
| add memory queue (ablation) | MATR · THUMOS14 / MUSES | average mAP | Improved |
| reduce segment size from 64 to 8 (ablation) | OAT-OSN vs MATR · THUMOS14 | average mAP: OAT-OSN 44.6 -> 25.8; MATR drop only 9.1%p | Hurt |
| add region classification for end prediction (ablation) | MATR end head · THUMOS14 | average mAP | Hurt |
| increase instance queries N from 10 to 30 (ablation) | MATR · THUMOS14 | average mAP: 49.5 -> 46.0 | Hurt |
| add anticipation region (Ta from 0 to 16) for predicting not-yet-ended actions (ablation) | MATR · THUMOS14 | average mAP: 40.2 -> 49.5 | Improved |
| NMS threshold far from 0.3 (0.1 or 0.7) (ablation) | MATR · THUMOS14 | average mAP: 49.5 -> 46.2 at 0.7 | Hurt |
| reduce segment length to 16 for faster inference (ablation) | MATR · THUMOS14 | average mAP and inference time: 49.5 -> 46.3 mAP; 167.1ms -> 53.8ms | Mixed |
| larger memory queue size (ablation) | MATR · THUMOS14 / MUSES | average mAP | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CAG-QIL — CAG-QIL: Context-Aware Actionness Grouping via Q Imitation Learning for Online Temporal Action Localization (ICCV 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- A Sliding Window Scheme for Online Temporal Action Localization (ECCV 2022 · similarity 0.92)
- TALLFormer: Temporal Action Localization with a Long-Memory Transformer (ECCV 2022 · similarity 0.90)
- Action Sensitivity Learning for Temporal Action Localization (ICCV 2023 · similarity 0.90)
- Temporal Action Localization in the Deep Learning Era: A Survey (TPAMI 2023 · similarity 0.90)
- Test-Time Zero-Shot Temporal Action Localization (CVPR 2024 · similarity 0.89)
- Energy vs. Noise: Towards Robust Temporal Action Localization in Open-World (AAAI 2025 · similarity 0.89)
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.
