Research map / Video / Action recognition
Long Short-Term Transformer for Online Action Detection
Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, Stefano Soatto
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-stage rather than one-stage memory compression (ablation) | LSTR · THUMOS'14, m_L=512s | mAP: 68.7 -> 69.5 | Improved |
| use two-stage compression for speed (ablation) | LSTR · THUMOS'14 runtime | FPS: 59.5 -> 91.6 | Improved |
| add LSTR encoder (learned long-term compression) before decoder (ablation) | LSTR · THUMOS'14, m_L=512s | mAP: 67.0 -> 69.5 | Improved |
| add LSTR decoder using short-term memory as queries (ablation) | LSTR · THUMOS'14, m_L=512s | mAP: 65.4 -> 69.5 | Improved |
| use learned adaptive compression instead of strided downsampling to 16 features (ablation) | LSTR · THUMOS'14 | mAP: 65.9 -> 69.5 | Improved |
| cross-attention instead of average pooling or concatenation for memory integration (ablation) | LSTR · THUMOS'14, m_L=512s | mAP: 68.7 -> 69.5 | Improved |
| transformer decoder units rather than self-attention encoder for the LSTR encoder (ablation) | LSTR · THUMOS'14, m_L=512s | mAP: 67.0 -> 69.5 | Improved |
| learned attention compression instead of a projection layer (ablation) | LSTR · THUMOS'14, m_L=512s | mAP: 66.9 -> 69.5 | Improved |
| separate long/short memory with encoder-decoder instead of single Transformer encoder on concatenated memory (ablation) | LSTR · THUMOS'14, m_L=512s | mAP: 66.5 -> 69.5 | Improved |
| downsample long-term memory with stride up to 4 (ablation) | LSTR · THUMOS'14 | mAP: 69.5 -> 69.5 | No clear effect |
| add GRU to cover history beyond memories (ablation) | LSTR · THUMOS'14 | mAP | Mixed |
| increase long-term memory length (ablation) | LSTR · THUMOS'14 | mAP: 66.6 -> 69.5 | Improved |
Closest papers by meaning
- OadTR: Online Action Detection With Transformers (ICCV 2021 · similarity 0.88)
- Lattice Long Short-Term Memory for Human Action Recognition (ICCV 2017 · similarity 0.87)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.86)
- Temporal Dynamic Graph LSTM for Action-Driven Video Object Detection (ICCV 2017 · similarity 0.86)
- Context-Enhanced Memory-Refined Transformer for Online Action Detection (CVPR 2025 · similarity 0.86)
- Scaling Action Detection: AdaTAD++ with Transformer-Enhanced Temporal-Spatial Adaptation (ICCV 2025 · similarity 0.86)
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.
