zht022 / eco-pytorch

PyTorch implementation for "ECO: Efficient Convolutional Network for Online Video Understanding", ECCV 2018

ECO-pytorch

• We provide the latest version of ECO-pytorch and pretrained models here.

• Many thanks to the author @mzolfaghari.

• If you have any questions, feel free to open a new issue in this repo.

• Pre-trained model for 2D-Net is provided by tsn-pytorch.

• Codes modified from tsn-pytorch.

PAPER INFO

"ECO: Efficient Convolutional Network for Online Video Understanding"
By Mohammadreza Zolfaghari, Kamaljeet Singh, Thomas Brox
paper link

Environment:

• Python 3.6.4
• PyTorch 0.3.1

Clone this repo

git clone https://github.com/zhang-can/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

[UCF101 - ECO - RGB] command:

python main.py ucf101 RGB <ucf101_rgb_train_list> <ucf101_rgb_val_list> \
        --arch ECO --num_segments 4 --gd 5 --lr 0.001 --lr_steps 30 60 --epochs 80 \
        -b 32 -i 1 -j 1 --dropout 0.8 --snapshot_pref ucf101_ECO --rgb_prefix img_ \
        --consensus_type identity --eval-freq 1