search

VISION-SJTU / RECCE

[CVPR2022] End-to-End Reconstruction-Classification Learning for Face Forgery Detection
MIT License
103 stars 14 forks source link

data pre-processing and dataset initialization #3

Closed MZMMSEC closed 1 year ago

MZMMSEC commented 2 years ago

Hi,

thanks for your sharing the codes! I have some questions about the data pre-processing and dataset initialization: 1. in the FaceForensics class in faceforensics.py, what is the '.pickle' file? and would you please provide this file? 2. how many frames per video do you extracted for detecting the face and further for training? especially how to balance the data number between the real videos and fake videos?(that is, extract the same number of frames first and then balance them in the training or extract the different number of frames at the begining for data balance?)

thanks~!

XJay18 commented 2 years ago

Hi, the pickle files just store the image paths and binary labels for experiments. You may use your own dataloder to load images.

We extract roughly the same number of frames per video for detecting the face and store the cropped faces. We sample about 12 frames per manipulated video for training, which results in 36,000 (=~12.5 frames/video×720 videos×4 manipulated techniques) fake images. The number of real images used for training is also 36,000 (=~50 frames/video×720 videos).

MZMMSEC commented 1 year ago

Hi, the pickle files just store the image paths and binary labels for experiments. You may use your own dataloder to load images.

We extract roughly the same number of frames per video for detecting the face and store the cropped faces. We sample about 12 frames per manipulated video for training, which results in 36,000 (=~12.5 frames/video×720 videos×4 manipulated techniques) fake images. The number of real images used for training is also 36,000 (=~50 frames/video×720 videos).

Hi, sorry to bother you again! Recently, I was doing some experiments and research by using your provided codes. However, when I use your data processing codes, mentioned in #1, to crop the faces from videos, I didn't get the same or similar results (e.g. face size) of extracted faces, as your provided processed datasets, mentioned in #2. As mentioned in many papers, a differerent face pre-processing will lead to a drop performance on the original model.

So, I would like to request for your face preprocessing codes with the detailed parameters settings.

Thanks!

XJay18 commented 1 year ago

Hi, unfortunately, I am not able to retrieve the specific parameter settings for FF++ preprocessing because it's been too long since the preprocessing task.

MZMMSEC commented 1 year ago

Hi, unfortunately, I am not able to retrieve the specific parameter settings for FF++ preprocessing because it's been too long since the preprocessing task.

thanks all the same!