This repo is the official implementation of the paper of GIF-Thumbnails: GIF Thumbnails: Attract More Clicks to Your Videos(AAAI-21).
- NVIDIA P40
- Python 3.7
- Pytorch 1.4
Videos are provided in the json annotated.json and unannotated.json. For annotated data, json is organized as:
{
$video_hash_key: {
"gifs": {
$gif_id1: [idx1, idx2, idx3], ...
$gif_idn: [idx1, idx2, idx3],
},
"video_url": $url,
"height": $height,
"width": $width,
"fps": $fps,
"frames": $frames,
}, ...
}
For unannotated data, videos are listed as:
{
$video_hash_key: {
"seconds": $seconds,
"video_url": $url
}
}
Download videos by you-get
you-get --format=flv720 https://www.bilibili.com/video/av18182135 -O 18182135.flv
Then we process the video to fps=16 and new_height=240 with ffmpeg
ffmpeg -i $InputVideo -r 16 -vf scale=$new_width:$new_height -qscale 0 $OutputVideo
where
Considering that some videos has been removed and are not available on bilibili now, we provide the feature of these videos extracted by 3D-ResNet-50.
Features are released on the BaiduDisk with code r9p1
.
For time-sync comments, we only release the number of comments in each second.
{
$video_hash_key: [cnt_1, cnt_2, ..., cnt_n],
...
}
Time-sync comments and video share the same video_hash_key, and
If you find this project is useful for your research, please cite:
@inproceedings{xu2021gif,
title={GIF Thumbnails: Attract More Clicks to Your Videos},
author={Xu, Yi and Bai, Fan and Shi, Yingxuan and Chen, Qiuyu and Gao, Longwen and Tian, Kai and Zhou, Shuigeng and Sun, Huyang},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={35},
number={4},
pages={3074--3082},
year={2021}
}