minhkhoi1026/face-spoofing-dection

FACE SWAP DETECTION BASED ON LANDMARK FOCUSING FEATURE

Installation

The instruction assume you are on a linux-based operating system. Using other OS will have to incompatible and need to modify.

Prerequisites:

• Conda: https://docs.conda.io/en/latest/miniconda.html
• Python 3.8.16

conda create -n fsd python=3.8.16

Run this command in your bash terminal:

bash install_dep.sh

Then, install the source package using the following command in root folder:

pip install .

Data

Data are available at Kaggle.

Training and testing images are stored in test and train folders respectively. The folder will be organised as such: videos

dataset_v1/
|-- videos/
    |-- deepfaker/
        |-- real/
        |-- fake/
    |-- roop/
        |-- real/
        |-- fake/
|-- imgs/
    |-- deepfaker/
        |-- landmark/
            |-- real/
            |-- fake/
        |-- face/
            |-- real/
            |-- fake/
    |-- roop/
        |-- landmark/
            |-- real/
            |-- fake/
        |-- face/
            |-- real/
            |-- fake/

Extract data

cmd:

$ python scripts/extract_all.py  --source <> --dest <> --sampling-ratio <> --extract-type <>

params:

• source: path of source data
• dest: path of destination image store
• sampling-ratio: specify a ratio x for frame sampling (0 < x <= 1)
• extract-type: choices in {all, frame, face}, default=all

For example:

$ python scripts/extract_all.py  --source ./data/ --dest ./data_extract/ --sampling-ratio 0.5 --extract-type all

If file structer of source folder '''./data/''' is as such:

    data
	    └── real
	    	    └── real1.mp4
					.
					.
	    └── fake
	    	    └── fake1.jpg
					.
					.

Then file structer of source folder '''./data_extract/''' is as such:

    data
		└── face_crop
					└── real
							└── real1_0000.png
								real1_0005.png
								.
								.
					└── fake
							└── fake1_0000.png
								fake1_0005.png
								.
								.
		└── landmark
					└── real
							└── real1_0000_landmark.png
								real1_0005_landmark.png
								.
								.
					└── fake
							└── fake1_0000_landmark.png
								fake1_0005_landmark.png
								.
								.

Split dataset to train/test/val

python scripts/dataset_split.py --src <> --type <[normal,variant]> --sampling-ratio <>

For example:

python scripts/dataset_split.py --src fsd-deepfakerapp/dfapp_plus_landmark/train/ --type variant --sampling-ratio 0.5

Train

Choose appropriate config in configs folder, then run this command:

python src/train.py -c configs/your_config.yml

Test

Choose appropriate config in configs folder, then run this command:

python src/test.py -c configs/your_config.yml

User need to adjust these below fields:

• dataset.args.test
• data_loader.test
• global.resume: checkpoint path
• metric [optional]: change it whether you want different metrics with training stage """

Demo application

This command will invoke new web application at localhost:20626:

streamlit run demo/lkt2.py --server.port 20626

Note: the first time the program process query, it need some warm-up time.

If you want to change the list of detector, you can edit the information of configs/detector_registry_config.yml.