Research map / Video / Action recognition
DynamoNet: Dynamic Action and Motion Network
Ali Diba, Vivek Sharma, Luc Van Gool, Rainer Stiefelhagen
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 frame prediction loss to train dynamic motion filters (ablation) | DynamoNet · UCF101 (trained on UCF101) | accuracy (%): 43.2 -> 50.2 | Improved |
| add dynamic motion filters with frame prediction to STC-ResNext101 (16 frames) | STC-ResNext101 · Kinetics-400 val | top-1 accuracy (%): 66.2 -> 67.6 | Improved |
| add dynamic motion filters to 3D ResNext101 (16 frames) | 3D ResNext101 · Kinetics-400 val | top-1 accuracy (%): 65.1 -> 66.3 | Improved |
| increase input from 32 to 64 frames in DynamoNet (STCnet) (ablation) | DynamoNet (STCnet) · Kinetics-400 val | top-1 accuracy (%): 69.9 -> 77.9 | Improved |
| add dynamic motion filters to STC-ResNext101 trained on half of Kinetics | STC-ResNext101 · Kinetics-400 val (half training data) | top-1 accuracy (%): 55.4 -> 63.6 | Improved |
| add dynamic motion filters to STC-ResNext101 trained on full Kinetics | STC-ResNext101 · Kinetics-400 val (full training data) | top-1 accuracy (%): 66.2 -> 67.6 | Improved |
| add self-supervised frame-prediction pre-training vs scratch STC-ResNet101 | STC-ResNet 101 · UCF101 (self-supervised pre-training, split-1) | accuracy (%): 56.7 -> 88.1 | Improved |
| add self-supervised frame-prediction pre-training vs scratch STC-ResNet101 | STC-ResNet 101 · HMDB51 (self-supervised pre-training, split-1) | accuracy (%): 30.8 -> 59.9 | Improved |
| add dynamic motion filters to STC-ResNext101 (UCF101, Kinetics pre-trained) | STC-ResNext101 · UCF101 (Kinetics pre-trained, 3 splits) | accuracy (%): 92.3 -> 96.6 | Improved |
| vary dynamic filter size away from 5x5 (3,4,6,7) (ablation) | DynamoNet · action classification | classification accuracy: drop of approx 2-3% | Hurt |
| add motion filters to predict frames for next-frame prediction vs Mathieu et al. | Mathieu et al. · UCF101 test videos (frame prediction) | SSIM: 0.92 -> 0.95 | Improved |
Later papers that compared with this method
1 later accepted paper reports DynamoNet (64) as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- Self-supervised Motion Representation via Scattering Local Motion Cues (ECCV 2020 · better on 0 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
- Spatiotemporal Residual Networks for Video Action Recognition (NeurIPS 2016 · similarity 0.89)
- Multi-level Motion Attention for Human Motion Prediction (IJCV 2021 · similarity 0.89)
- Video Modeling With Correlation Networks (CVPR 2020 · similarity 0.89)
- Deep Insights into Convolutional Networks for Video Recognition (IJCV 2019 · similarity 0.88)
- Improving Video Model Transfer With Dynamic Representation Learning (CVPR 2022 · similarity 0.88)
- Lattice Long Short-Term Memory for Human Action Recognition (ICCV 2017 · 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.
