Research map / Video / Action recognition
TEA: Temporal Excitation and Aggregation for Action Recognition
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 MTA hierarchical temporal aggregation (ablation) | (2+1)D ResNet · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 46.0 -> 47.5 | Helped |
| add spatial Res2Net-style sub-convolutions without temporal receptive-field expansion (ablation) | (2+1)D ResNet · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 46.0%/46.2% | No clear effect |
| add ME motion encoding and channel excitation (ablation) | (2+1)D ResNet · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 46.0 -> 48.4 | Helped |
| replace ME with per-frame SE excitation (ablation) | (2+1)D ResNet · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 46.0 -> 46.5 | No clear effect |
| remove the residual connection from ME (ablation) | ME · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 48.4 -> 47.2 | Hurt |
| replace motion-based excitation with STM-style feature addition | ME · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 48.4 vs 47.5 | Hurt |
| add MTA to the ME model (ablation) | ME · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 48.4 -> 48.9 | Helped |
| use general randomly initialized temporal convolutions (ablation) | (2+1)D ResNet · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 43.4 -> 23.5 | Hurt |
| use randomly initialized channel-wise temporal convolutions (ablation) | (2+1)D ResNet · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 43.4 -> 43.6 | No clear effect |
| initialize learnable channel-wise temporal convolutions as part shift (ablation) | (2+1)D ResNet · Something-Something V1 val, 8×1×1 | Top-1 accuracy: 43.4 -> 46.0 | Helped |
| add a 3×3 transformation convolution on features at t+1 in ME (ablation) | TEA · Something-Something V1 val | Top-1 accuracy: 48.4% -> 48.9% | Helped |
| transform both t and t+1 features instead of only t+1 (ablation) | ME · Something-Something V1 val | Top-1 accuracy | No clear effect |
Closest papers by meaning
- Temporal Convolutional Networks for Action Segmentation and Detection (CVPR 2017 · similarity 0.89)
- ACTION-Net: Multipath Excitation for Action Recognition (CVPR 2021 · similarity 0.89)
- MS-TCN: Multi-Stage Temporal Convolutional Network for Action Segmentation (CVPR 2019 · similarity 0.88)
- Shrinking Temporal Attention in Transformers for Video Action Recognition (AAAI 2022 · similarity 0.88)
- End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames (CVPR 2024 · similarity 0.87)
- Taylor Videos for Action Recognition (ICML 2024 · similarity 0.87)
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.