Skip to content

03-preprocessing.md

Latest commit

 

History

History
82 lines (69 loc) · 2.68 KB

03-preprocessing.md

File metadata and controls

82 lines (69 loc) · 2.68 KB

SCRUM-35: Preprocessing Pipeline for XceptionNet

Dataset

Storage Location

  • Path: AI-Video-Detection/backend/dataset/AIGVDBench
  • Access command:
python3 -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='AIGVDBench/AIGVDBench', repo_type='dataset', local_dir='./AIGVDBench')"

Data Split

  • Train/Val/Test ratio: 70/15/15
  • Counts:
    • Train: 280000 videos
    • Val: 60000 videos
    • Test: 60000 videos
  • Method: Random split with seed 42; each unique real video and its derivatives remain in the same split.

Preprocessing Steps

  1. Initialize MTCNN based on input arguments.
  2. Use OpenCV to split video into RGB frames.
  3. Pass frames into MTCNN to detect faces.
  4. Generate a data list containing tagged face images.

Code Changes

  • Updated backend/main.py
  • Updated backend/handlers/video_handler.py
  • Updated backend/preprocessing/video_processor.py
  • Added backend/models/DeepFake-EfficientNet/extract_face_efficientNet.sh
  • Added backend/models/DeepFake-EfficientNet/scripts/test.py
  • Modified backend/models/DeepFake-EfficientNet/scripts/train.py
  • Added backend/dataset/AIGVDBench/AIGVDBench/
  • Added backend/dataset/AIGVDBench/AIGVDBench/split_videos_left_out_easyAnimate.py
  • Added backend/dataset/AIGVDBench/AIGVDBench/split_videos_standard_split.py

How to Run

# Step 1: Download/locate dataset
# Expected path: AI-Video-Detection/backend/dataset/AIGVDBench/AIGVDBench/

# Step 2: Split the dataset
cd backend/dataset/AIGVDBench/AIGVDBench
python3 split_videos_standard_split.py
# or
python3 split_videos_left_out_easyAnimate.py

# Step 3: Run preprocessing
cd backend/models/DeepFake-EfficientNet
./extract_face_efficientNet.sh

# Step 4: Generate data list
cd backend/models/XceptionNet-Detector/Deepfake-Detection
./datalist.sh

Verification

  • Ran preprocessing end-to-end
  • Verified split counts match expectations
  • Verified processed data looks correct (sample check)
  • No dataset files committed
  • Documentation updated

Notes

Note: The preprocessing pipeline of XceptionNet is basically same as the one of EfficientNet. To avoid repeating the work, XceptionNet uses the same preprocessing pipeline as EfficientNet.

Results: For AIGVDBench dataset, XceptionNet got the following: epoch train loss: 0.0000 Acc: 0.9996 epoch val loss: 0.0061 Acc: 0.9492 Best val Acc: 0.9507 Test Acc: 0.9505 For FaceForensics++ dataset, XceptionNet got the following: epoch train loss: 0.0000 Acc: 0.9992 epoch val loss: 0.0104 Acc: 0.9349 Best val Acc: 0.9364 Test Acc: 0.9016