Research map / Video / Action recognition
STMixer: A One-Stage Sparse Action Detector
Tao Wu, Mengqi Cao, Ziteng Gao, Gangshan Wu, Limin Wang
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 adaptive sampling rather than fixed-grid sampling (ablation) | STMixer · AVA v2.2 val | mAP: 21.9 -> 23.1 | Improved |
| Use a unified 4D feature space rather than separate key-frame classification and Res5 localization features (ablation) | STMixer · AVA v2.2 val | mAP: 22.8 -> 23.1 | Improved |
| Use simple lateral convolution instead of full FPN (ablation) | STMixer 4D feature construction · AVA v2.2 val | mAP: 23.1 -> 22.9 | No clear effect |
| Move sampling points across frames instead of copying them (ablation) | STMixer adaptive sampling · AVA v2.2 val | mAP: 23.1 -> 23.3 | Improved |
| Use query-guided adaptive mixing rather than fixed-parameter mixing (ablation) | STMixer feature mixing · AVA v2.2 val | mAP: 22.5 -> 23.1 | Improved |
| Use parallel dual-branch spatiotemporal mixing rather than sequential spatiotemporal mixing (ablation) | STMixer feature mixing · AVA v2.2 val | mAP: 22.6 -> 23.1 | Improved |
| Use both spatial and temporal mixing rather than spatial mixing alone (ablation) | STMixer feature mixing · AVA v2.2 val | mAP: 22.4 -> 23.1 | Improved |
| Use both spatial and temporal mixing rather than temporal mixing alone (ablation) | STMixer feature mixing · AVA v2.2 val | mAP: 22.1 -> 23.1 | Improved |
| Use 32 sampling points per frame (ablation) | STMixer adaptive sampling · AVA v2.2 val | mAP: 22.4 (8 points), 22.5 (16 points), 23.1 (32 points), 23.0 (48 points), 22.5 (64 points) | Mixed |
| Use 100 queries (ablation) | STMixer decoder · AVA v2.2 val | mAP: 22.1 (15 queries), 22.9 (50 queries), 23.1 (100 queries), 22.5 (150 queries) | Mixed |
| Use 6 ASAM modules (ablation) | STMixer decoder · AVA v2.2 val | mAP: 18.4 (1 module), 22.5 (3 modules), 23.1 (6 modules), 22.6 (9 modules) | Mixed |
| Add long-term query support | STMixer with SlowFast-R101 · AVA v2.2 | mAP: 30.9 -> 32.9 | Improved |
Later papers that compared with this method
2 later accepted papers report STMixer as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Classification Matters: Improving Video Action Detection with Class-Specific Attention (ECCV 2024 · better on 3 of 3 shared results)
- Efficient Video Action Detection with Token Dropout and Context Refinement (ICCV 2023 · better on 2 of 2 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
- Watch Only Once: An End-to-End Video Action Detection Framework (ICCV 2021 · similarity 0.88)
- SV-TAD: Native Sparse Convolutions for Efficient Temporal Action Detection (ECCV 2026 · similarity 0.87)
- PointTAD: Multi-Label Temporal Action Detection with Learnable Query Points (NeurIPS 2022 · similarity 0.86)
- Learning To Refactor Action and Co-Occurrence Features for Temporal Action Localization (CVPR 2022 · similarity 0.86)
- Dual DETRs for Multi-Label Temporal Action Detection (CVPR 2024 · similarity 0.86)
- TrAction: Action Recognition with Sparse Trajectories (NeurIPS 2026 · 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.
