Research map / Video / Action recognition
AdaFuse: Adaptive Temporal Fusion Network for Efficient Action Recognition
Yue Meng, Rameswar Panda, Chung-Ching Lin, Prasanna Sattigeri, Leonid Karlinsky, Kate Saenko, Aude Oliva, Rogerio Feris
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 adaptive temporal fusion to ResNet18 TSN | TSN with ResNet18 · Something-Something V1 | Top1 accuracy: 14.8 -> 36.9 | Improved |
| Add adaptive temporal fusion to ResNet18 TSN | TSN with ResNet18 · Something-Something V2 | Top1 accuracy: 27.3 -> 50.5 | Improved |
| Add adaptive temporal fusion to ResNet18 TSN | TSN with ResNet18 · Jester | Top1 accuracy: 82.6 -> 93.7 | Improved |
| Add adaptive temporal fusion to ResNet18 TSN | TSN with ResNet18 · Mini-Kinetics | Top1 accuracy: 64.6 -> 67.5 | Improved |
| Add feature reuse to learned keep/skip policies (ablation) | Ada. Skip with ResNet18 · Something-Something V1 | Top1 accuracy: 9.5 -> 36.9 | Improved |
| Add channel skipping to learned keep/reuse policies (ablation) | Ada. Reuse with ResNet18 · Something-Something V1 | FLOPS: 13.8G -> 10.3G | Improved |
| Add channel skipping to learned keep/reuse policies (ablation) | Ada. Reuse with ResNet18 · Something-Something V1 | Top1 accuracy: 36.3 -> 36.9 | Improved |
| Replace random channel policy with an adaptive learned policy (ablation) | Temporal fusion with ResNet18 · Something-Something V1 | Top1 accuracy: 27.5 -> 36.9 | Improved |
| Restrict adaptive fusion to the last three TSM blocks rather than all blocks (ablation) | TSM with ResNet50 · Something-Something V1 | Top1 accuracy: 44.9 -> 46.8 | Improved |
| Increase policy hidden units from 1024 to 4096 (ablation) | AdaFuse-TSM-Last · Something-Something V2 | Top1 accuracy: 59.40 -> 60.00 | Improved |
| Increase efficiency weight lambda from 0.050 to 0.100 (ablation) | AdaFuse-TSM-Last with 1024 hidden units · Something-Something V2 | FLOPS: 31.53G -> 31.04G | Improved |
| Increase efficiency weight lambda from 0.050 to 0.100 (ablation) | AdaFuse-TSM-Last with 1024 hidden units · Something-Something V2 | Top1 accuracy: 59.71 -> 59.40 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AR-Net — AR-Net: Adaptive Frame Resolution for Efficient Action Recognition (ECCV 2020) · 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
- Convolutional Two-Stream Network Fusion for Video Action Recognition (CVPR 2016 · similarity 0.88)
- Memory-Augmented Temporal Dynamic Learning for Action Recognition (AAAI 2019 · similarity 0.87)
- EAN: Event Adaptive Network for Enhanced Action Recognition (IJCV 2022 · similarity 0.87)
- Action Recognition With Coarse-to-Fine Deep Feature Integration and Asynchronous Fusion (AAAI 2018 · similarity 0.87)
- TEINet: Towards an Efficient Architecture for Video Recognition (AAAI 2020 · similarity 0.87)
- Spatiotemporal Pyramid Network for Video Action Recognition (CVPR 2017 · 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.
