Anti-deepfake

Preprocess

• Usage

1. Generate frames, will save the frames under ../dataset/frames/

   cd ./tools/
   python generate_frames.py path/to/videos --cpu_num 48

2. Generate face patches, will save the patches under ../dataset/patches/ and will save a csv file patch_image_statics.csv.

   cd ./tools/
   python generate_patches.py path/to/frames/

   patch_image_statics.csv contains PatchName and Score:

   PatchName	Score
   uaspniazcl_002_face_0.jpg	0.9996267557144165
   uaspniazcl_003_face_0.jpg	0.9994097948074341
   uaspniazcl_001_face_0.jpg	0.9971966743469238

   Remember to collect this file after a batch of computation.

3. Generate training label csv.

   cd ./tools/
   python anno_parse.py path/to/patch/file path/to/jsonfile --saving_path path/to/save

• Name rules

1. Frame images:

   VideoName_FrameId.jpg eg. rmufsuogzn_006.jpg, vxawghqzyf_007.jpg.

2. Patch images:

   VideoName_FrameId_face_FaceId.jpg eg. ucthmsajay_008_face_0.jpg, zzlsynxeff_008_face_0.jpg

Training

Train the Exception Net. Need to change some path in the code.

cd ./tools/
python train_net.py

While using single GPU training please use:

cd ./tools/
CUDA_VISIBLE_DEVICES=1,2,3 python train_net.py

After training, the best model and checkpoints will saved under ./saved_models/model.pth

Inference

Do inference on video level.

cd ./tools/
python infer_net.py

After inference, the result will be saved under ./dataset/submission.csv

TODO List

• XceptionNet
• Basic training logic
• Weighted training
• Inference
• Data augmentation
• Better face detection model/ finetune face detection model