HDTF

Flow-guided One-shot Talking Face Generation with a High-resolution Audio-visual Dataset paper supplementary demo video

Details of HDTF dataset

./HDTF_dataset consists of youtube video url, video resolution (in our method, may not be the best resolution), time stamps of talking face, facial region (in the our method) and the zoom scale of the cropped window.

xx_video_url.txt:

format:     video name | video youtube url

xx_resolution.txt:

format:    video name | resolution(in our method)

xx_annotion_time.txt:

format:    video name | time stamps of clip1 | time stamps of clip2 | time stamps of clip3....

xx_crop_wh.txt:

format:    video name+clip index | min_width | width |  min_height | height (in our method)

xx_crop_ratio.txt:

format:    video name+clip index | window zoom scale

Processing of HDTF dataset

When using HDTF dataset,

• We provide video and url in xx_video_url.txt. (the highest definition of videos are 1080P or 720P). Transform video into .mp4 format and transform interlaced video to progressive video as well.

• We split long original video into talking head clips with time stamps in xx_annotion_time.txt. Name the splitted clip as video name_clip index.mp4. For example, split the video Radio11.mp4 00:30-01:00 01:30-02:30 into Radio11_0.mp4 and Radio11_1.mp4 .

• Our work does not always download videos with the best resolution, so we provide two cropping methods. Thanks @universome and @Feii Yin for pointing out this problem!

  1. Download the video with reference resulotion in xx_resolution.txt and crop the facial region with fixed window size in xx_crop_wh.txt. (This method is as same as ours, but the downloaded video may not be the best resolution).
  2. First, download the video with best resulotion. Then, detect the facial landmark in the splitted talking head clips and count the square window of the face, specifically, count the facial region in each frame and merge all regions into one square range. Next, enlarge the window size with xx_crop_ratio.txt. Finally, crop the facial region.

• We resize all cropped videos into 512 x 512 resolution.

The HDTF dataset is available to download under a Creative Commons Attribution 4.0 International License. Thanks @universome for provding the the script of data processing, pls visit here for more details. If you face any problems when processing HDTF, pls contact me.

Inference code

code of audio-to-animation

coming soon......

code of constructing approximate dense flow

The code is in ./code_constructing_Fapp, pls visit here for more details.

code of animation-to-video module

The code is in ./code_animation2video, pls visit here for more details.

code of reproducing other works

coming soon......

Reference

if you use HDTF, pls reference

@inproceedings{zhang2021flow,
  title={Flow-Guided One-Shot Talking Face Generation With a High-Resolution Audio-Visual Dataset},
  author={Zhang, Zhimeng and Li, Lincheng and Ding, Yu and Fan, Changjie},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={3661--3670},
  year={2021}
}