We release the code of the DSANet (Dynamic Segment Aggregation Network). We introduce the DSA module to capture relationship among snippets for video-level representation learning. Equipped with DSA modules, the top-1 accuracy of I3D ResNet-50 is improved to 78.2% on Kinetics-400.

The core code to implement the Dynamic Segment Aggregation Module is codes/models/modules_maker/DSA.py.

[July 7, 2021] We release the core code of DSANet.

[July 3, 2021] DSANet has been accepted by ACMMM 2021.

• Prerequisites
• Data Preparation
• Model Zoo
• Testing
• Training

All dependencies can be installed using pip:

python -m pip install -r requirements.txt

Our experiments run on Python 3.7 and PyTorch 1.5. Other versions should work but are not tested.

• Download ImageNet pre-trained models for offline environment

cd pretrained
sh download_imgnet.sh

• Download K400 pre-trained models for inference

TODO

We follow the same data process with MVFNet for data preparation.

TODO

bash dist_test_recognizer.sh CONFIG_PATH CHECKPOINT_PATH 8 

This implementation supports multi-gpu, DistributedDataParallel training, which is faster and simpler.

• For example, to train DSANet with 8 gpus, you can run:

bash dist_train_recognizer.sh configs/kinetics/r50_e100.py 8

We especially thank the contributors of the MVFNet and mmaction codebase for providing helpful code.

This repository is released under the Apache-2.0. license as found in the LICENSE file.

MVFNet: Multi-View Fusion Network for Efficient Video Recognition, AAAI2021 Paper | Code

If you think our work is useful, please feel free to cite our paper 😆 :

@inproceedings{wu2021dsanet,
  title={DSANet: Dynamic Segment Aggregation Network for Video-Level Representation Learning},
  author={Wu, Wenhao and Zhao, Yuxiang and Xu, Yanwu and Tan, Xiao and He, Dongliang and Zou, Zhikang and Ye, Jin and Li, Yingying and Yao, Mingde and Dong, Zichao and others},
  booktitle = {ACMMM},
  year={2021}
}

For any question, please file an issue or contact

Wenhao Wu: [email protected]
Yuxiang Zhao: [email protected]