search

CHELSEA234 / Multi-domain-learning-FAS

The multi-domain FAS work with SiW-Mv2 dataset (ECCV 2022 oral)
http://cvlab.cse.msu.edu/siwm-v2-dataset.html
66 stars 9 forks source link

about data preprocessing #4

Closed woody-panda closed 1 year ago

woody-panda commented 1 year ago

Thanks for sharing the code. I'm new to FAS and would be very grateful if you could give me some data preprocessing details.

  1. According to the paper, the FAS model and framework are frame-based, right? In preprocessing, is it necessary to extract frames from the video? Then crop out the face from each frame? How many frames are extract from each video?
CHELSEA234 commented 1 year ago

Thanks to your comments. :) Yes. Most FAS models are frame-based, except for a few works that takes videos (e.g., a few consecutive frames) as input. In my opinion, you have to extract frames given videos, and this is just one line of code using OpenCV; Cropping the face with bounding box, you can go with the off-the-shelf tools, such as RetinaFace and MTCNN.

tangwudu commented 1 year ago

You will just use original cropped face or any padding based on face bounding box?