load_videos.py - Can not load video UW1c9E8nfxQ, broken link about face-vid2vid

[zhengkw18]

Some video links may be invalid since they come from Youtube. It seems that you can't reach any of the links. You should check whether you have access to those Youtube videos, for example visit the sites directly.

[brbernardo90]

@zhengkw18 Thanks for your quick reply.

You are right, but I tried to access the videos directly and some are possible and some no longer exist. I tried to solve it but i couldn't. Anyway, I have my own dataset and I am trying understand how to organize my dataset to adapt to your model. Can you help me?

Each lesson of my dataset is composed by:

• The file keypoints txt (driving, array shape = [168,2]);
• A folder with the frames png (target video, img size = [256,256])
  .
  └── facedataset/
  ├── test_img/
  │ ├── video_name_1/
  │ │ ├── 00001.png
  │ │ ├── 00002.png
  │ │ ├── 00003.png
  │ │ └── ...
  │ ├── video_name_2
  │ └── video_name_3
  ├── test_keypoints/
  │ ├── video_name_1/
  │ │ ├── 00001.txt
  │ │ ├── 00002.txt
  │ │ ├── 00003.txt
  │ │ └── ...
  │ ├── video_name_2
  │ └── video_name_3
  ├── train_img
  └── train_keypoints

[zhengkw18]

You only need to prepare folders with the frames png, because we extract the headpose using off-the-shelf model, and learn a 3D keypoint estimator from scratch. The folder structure is:
├──datasets/vox/train/
│ ├── id1#video1/
│ │ ├── 00001.png
│ │ ├── 00002.png
│ │ ├── 00003.png
│ │ └── ...
│ ├── id1#video2
│ └── ...
│ ├── id2#video3
│ └── ...

Note that every folder contains frames png of a segment of a talking head video, and the folder should be named as id#video. The "#" is necessary, because we enable "id sampling". We'll view the name part before "#" as the person id, and uniformly sample the id, see dataset.py.