by Yang Feng, Lin Ma, Wei Liu, Tong Zhang, Jiebo Luo
Video Re-localization aims to accurately localize a segment in a reference video such that the segment semantically corresponds to a query video. For more details, please refer to our paper.
@InProceedings{feng2018video,
author = {Feng, Yang and Ma, Lin and Liu, Wei and Zhang, Tong and Luo,
Jiebo},
title = {Video Re-localization},
booktitle = {ECCV},
year = {2018}
}
sudo apt install python-opencv
pip install tf-nightly-gpu==1.11.0.dev20180821
-
Download the ActivityNet features at link. You will get activitynet_v1-3.part-00 to activitynet_v1-3.part-05.
-
Merge and unzip the files. You'll get sub_activitynet_v1-3.c3d.hdf5.
cat activitynet_v1-3.part-0* > temp.zip unzip temp.zip
-
Get the code and split the features.
git clone https://github.com/fengyang0317/video_reloc.git cd video_reloc ln -s /path/to/sub_activitynet_v1-3.c3d.hdf5 data/ wget "http://ec2-52-11-11-89.us-west-2.compute.amazonaws.com/files/\ activity_net.v1-3.min.json" -P data/ python split_feat.py
-
[Optional] Download the all the videos into data/videos and get the number of frames in each video.
python get_frame_num.py
-
Generate the dataset json.
python create_dataset.py
-
Train the model.
python match.py --data_dir data
-
Eval the model. The results may be slightly different from the values reported in the paper.
python eval.py --data_dir data --ckpt_path saving/model.ckpt-(best val ckpt)