Research map / Video / Action recognition
Backtrace Mamba: Reviving Critical Temporal Contexts via Hierarchical Memory Compression for Online Action Detection
Su Yan, Jiahua Li, Kun Wei, Cheng Deng
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 hierarchical memory mechanism to Mamba baseline (ablation) | Mamba OAD baseline · FineAction | mAP: 32.4 -> 36.8 | Improved |
| add temporal soft pruning on top of HMM (ablation) | MOAD · FineAction | mAP: 36.8 -> 41.2 | Improved |
| add temporal soft pruning on top of HMM (ablation) | MOAD · FineAction | FLOPs (G): 16.8 -> 10.2 | Improved |
| lower novelty threshold theta from 0.5 to 0.2 (ablation) | MOAD · THUMOS'14 (Kinetics features) | mAP: 70.4 -> 74.3 | Improved |
| increase PCA dimension from 32 to 48 (ablation) | MOAD · THUMOS'14 (Kinetics features) | mAP: 71.2 -> 74.3 | Improved |
| increase PCA dimension from 48 to 64 (ablation) | MOAD · THUMOS'14 (Kinetics features) | mAP: 74.3 -> 74.4 | No clear effect |
| increase codebook size from 128 to 192 (ablation) | MOAD · THUMOS'14 (Kinetics features) | mAP: 72.1 -> 74.3 | Improved |
| increase codebook size from 192 to 256 (ablation) | MOAD · THUMOS'14 (Kinetics features) | mAP: 74.3 -> 74.5 | No clear effect |
| merge two channels by sum instead of average (ablation) | MOAD · THUMOS'14 (Kinetics features) | mAP: 72.8 -> 74.3 | Improved |
| use both channels instead of one channel (ablation) | MOAD · THUMOS'14 (Kinetics features) | mAP: 70.6 -> 74.3 | Improved |
| replace transformer/RNN with Mamba-based MOAD | OAD · FineAction | mAP: 37.1 -> 41.2 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- MiniROAD — MiniROAD: Minimal RNN Framework for Online Action Detection (ICCV 2023) · better on 1 of 1 shared result
- OadTR — OadTR: Online Action Detection With Transformers (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
- Context-Enhanced Memory-Refined Transformer for Online Action Detection (CVPR 2025 · similarity 0.90)
- Mamba Only Glances Once (MOGO): A Lightweight Framework for Efficient Video Action Detection (NeurIPS 2025 · similarity 0.89)
- MS-Temba: Multi-Scale Temporal Mamba for Understanding Long Untrimmed Videos (CVPR 2026 · similarity 0.88)
- OnlineTAS: An Online Baseline for Temporal Action Segmentation (NeurIPS 2024 · similarity 0.86)
- Does Video-Text Pretraining Help Open-Vocabulary Online Action Detection? (NeurIPS 2024 · similarity 0.86)
- Temporal Action Detection Model Compression by Progressive Block Drop (CVPR 2025 · similarity 0.85)
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.
