Research map / Video / Action recognition
DyFADet: Dynamic Feature Aggregation for Temporal Action Detection
Le Yang, Ziwei Zheng, Yizeng Han, Hao Cheng, Shiji Song, Gao Huang, Fan Li
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 |
|---|---|---|---|
| replace conv encoder with deformable conv (ablation) | TAD baseline · THUMOS14 | avg mAP: 62.1 -> 66.1 | Improved |
| replace conv encoder with squeeze-and-excitation (ablation) | TAD baseline · THUMOS14 | avg mAP: 62.1 -> 63.4 | Improved |
| replace conv with DFA_Conv in encoder (ablation) | TAD baseline · THUMOS14 | avg mAP: 62.1 -> 66.8 | Improved |
| use DynE layer encoder instead of DFA_Conv (ablation) | DyFADet · THUMOS14 | avg mAP: 66.8 -> 67.8 | Improved |
| add dynamic multi-scale head (DyHead) on top of DynE (ablation) | DyFADet · THUMOS14 | avg mAP: 67.8 -> 69.2 | Improved |
| connect scales in head with plain convolution instead of DyHead (ablation) | DyFADet · THUMOS14 | avg mAP: 67.8 -> 68.0 | No clear effect |
| add DyHead to a plain conv encoder (ablation) | TAD baseline · THUMOS14 | avg mAP: 62.1 -> 65.8 | Improved |
| expansion factor w=5 vs 3 or 6 (ablation) | DyFADet · THUMOS14 | avg mAP: w=3: 68.2, w=5: 69.2, w=6: 67.0 | Mixed |
| head depth D=3 vs 2 or 4 (ablation) | DyFADet · THUMOS14 | avg mAP: D=2: 67.4, D=3: 69.2, D=4: 67.1 | Mixed |
| use C-formation vs K or CK formation for mask (ablation) | DyFADet · THUMOS14 | avg mAP: 68.4 (C, CK) vs 69.2 (K) | No clear effect |
| add optical flow features to VideoMAE V2-g | DyFADet · THUMOS14 | avg mAP: 70.5 -> 71.1 | Improved |
| use DynE encoder without DyHead instead of self-attention | ActionFormer · THUMOS14 latency | GPU latency (ms): 68.4 -> 44.7 | Improved |
Later papers that compared with this method
3 later accepted papers report DyFADet as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Decompose and Transfer: CoT-Prompting Enhanced Alignment for Open-Vocabulary Temporal Action Detection (CVPR 2026 · better on 2 of 2 shared results)
- Light but Sharp: SlimSTAD for Real-Time Action Detection from Sensor Data (AAAI 2026 · better on 2 of 2 shared results)
- DeFB: Decomposed Feature Learning for Real-Time Multi-Person Eyeblink Detection in Untrimmed In-the-Wild Videos (AAAI 2026 · better on 1 of 1 shared result)
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.
- ActionFormer — ActionFormer: Localizing Moments of Actions with Transformers (ECCV 2022) · better on 6 of 6 shared results
- TriDet — TriDet: Temporal Action Detection With Relative Boundary Modeling (CVPR 2023) · better on 4 of 5 shared results
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
- TEA: Temporal Excitation and Aggregation for Action Recognition (CVPR 2020 · similarity 0.87)
- End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames (CVPR 2024 · similarity 0.86)
- Dual DETRs for Multi-Label Temporal Action Detection (CVPR 2024 · similarity 0.86)
- TE-TAD: Towards Full End-to-End Temporal Action Detection via Time-Aligned Coordinate Expression (CVPR 2024 · similarity 0.85)
- Benchmarking the Robustness of Temporal Action Detection Models Against Temporal Corruptions (CVPR 2024 · similarity 0.84)
- DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer (CVPR 2025 · similarity 0.84)
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.
