This is a fork of Can Zhang's PyTorch implementation for the paper:

" ECO: Efficient Convolutional Network for Online Video Understanding, European Conference on Computer Vision (ECCV), 2018." By Mohammadreza Zolfaghari, Kamaljeet Singh, Thomas Brox

Update

• 2019.3.05: This is a major update. This update is more robust and we solved some problems in the previous version such as iter_size and ECO Full model definiation. Updating the training procedure (main.py) and providing the pretrained models for ECOLite and ECOFull. Please let us know if you found any problem or had suggestions to improve the code.

NOTE

• Trained models on Kinetics dataset for ECO Lite and C3D are provided.
• Pre-trained model for 2D-Net is provided by tsn-pytorch.
• Stay tuned for more updates

Environment:

• Python 3.5.2
• PyTorch 0.4.1
• TorchVison: 0.2.1

Clone this repo

git clone https://github.com/mzolfaghari/ECO-pytorch

Generate dataset lists

python gen_dataset_lists.py <ucf101/something> <dataset_frames_root_path>

e.g. python gen_dataset_lists.py something ~/dataset/20bn-something-something-v1/

The dataset should be organized as:
<dataset_frames_root_path>/<video_name>/<frame_images>

Training

1. Download the initialization and trained models:

      ECO-Lite pretrained model on Kinetics: https://drive.google.com/open?id=1XNIq7byciKgrn011jLBggd2g79jKX4uD
      ECO-Full pretrained model on Kinetics: https://drive.google.com/open?id=1ATuN_KctsbFAbcNgWDlETZVsy2vhxZay

Othe models:

      sh models/download_models.sh

• If you can not access Google Drive, please download the pretrained models from BaiduYun, and put them in "models" folder.

2. Command for training ECO Lite model:

    ./scripts/run_ECOLite_kinetics.sh local

3. For training C3D network use the following command:

    ./scripts/run_c3dres_kinetics.sh local

4. For finetuning on UCF101 use the following command:

    sh run_demo_ECO_Lite.sh local 
    or 
    sh run_demo_ECO_Full.sh local

NOTE

• If you want to train your model from scratch change the config as following:

    --pretrained_parts scratch

• configurations explained in "opts.py"

TODO

1. Trained models on other datasets

Citation

If you use this code or ideas from the paper for your research, please cite our paper:

@inproceedings{ECO_eccv18,
author={Mohammadreza Zolfaghari and
               Kamaljeet Singh and
               Thomas Brox},
title={{ECO:} Efficient Convolutional Network for Online Video Understanding},	       
booktitle={ECCV},
year={2018}
}

Contact

Mohammadreza Zolfaghari, Can Zhang

Questions can also be left as issues in the repository. We will be happy to answer them.