Research map / Video / Action recognition
Depthwise Spatio-Temporal STFT Convolutional Neural Networks for Human Action Recognition
Sudhakar Kumawat, Manisha Verma, Yuta Nakashima, Shanmuganathan Raman
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 fixed STFT kernel instead of trainable depthwise temporal conv (ablation) | T-STFT · Something-Something v1 (64 frames) | top-1 accuracy: 46.75 -> 50.63 | Improved |
| apply STFT in depthwise fashion instead of ReLPV blocks (ablation) | ST-STFT · Something-Something v1 (64 frames) | top-1 accuracy: 43.01 -> 48.23 | Improved |
| use LeakyReLU instead of ReLU (ablation) | ST-STFT · Something-Something v1 (16 frames) | top-1 accuracy: 42.62 -> 43.78 | Improved |
| use LeakyReLU instead of SELU (ablation) | ST-STFT · Something-Something v1 (16 frames) | top-1 accuracy: 32.83 -> 43.78 | Improved |
| use LeakyReLU instead of ELU (ablation) | ST-STFT · Something-Something v1 (16 frames) | top-1 accuracy: 32.39 -> 43.78 | Improved |
| use temporal-only STFT instead of spatio-temporal STFT | STFT network · Something-Something v1 (64 frames) | top-1 accuracy: 48.23 -> 50.63 | Improved |
| use temporal STFT instead of trainable depthwise temporal (S-STFT) | STFT network · Something-Something v1 (64 frames) | top-1 accuracy: 47.12 -> 50.63 | Improved |
| increase input frames from 16 to 64 | T-STFT · Something-Something v1 (from scratch) | top-1 accuracy: 46.74 -> 50.63 | Improved |
| add Kinetics pre-training | T-STFT · Something-Something v1 (64 frames) | top-1 accuracy: 50.63 -> 52.42 | Improved |
| add Kinetics pre-training | T-STFT · Something-Something v2 (64 frames) | top-1 accuracy: 63.05 -> 64.66 | Improved |
| add Kinetics pre-training | T-STFT · Jester (16 frames) | top-1 accuracy: 96.85 -> 96.94 | No clear effect |
| train from scratch without ImageNet pretraining | T-STFT · Kinetics-400 | top-1 accuracy: 73.7 -> 61.1 | Hurt |
Closest papers by meaning
- StNet: Local and Global Spatial-Temporal Modeling for Action Recognition (AAAI 2019 · similarity 0.89)
- STM: SpatioTemporal and Motion Encoding for Action Recognition (ICCV 2019 · similarity 0.89)
- Spatio-Temporal Channel Correlation Networks for Action Classification (ECCV 2018 · similarity 0.89)
- A Closer Look at Spatiotemporal Convolutions for Action Recognition (CVPR 2018 · similarity 0.88)
- Learning SpatioTemporal and Motion Features in a Unified 2D Network for Action Recognition (TPAMI 2022 · similarity 0.88)
- Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video Classification (ECCV 2018 · similarity 0.88)
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.
