Liveness Detection Baseline

This repository provides some scripts to easily train and test some timm models for Zalo AI Challenge 2022: Liveness Detection.

Approach:

• Simple image classification problem
• The module consists of a pretrained model from timm with an output size of 1 and a sigmoid layer
• Training and testing data are created by randomly sampling 10 frames from each video
• When training, each frame is fed into the model irrespective of which video it comes from
• When testing, the predictions from all 10 frames of each video are averaged to produce a final prediction

Dependencies

pip3 install -r requirements.txt

Data

Download the training and public test data from the competition portal to folder data and extract them. The structure of the folder will be:

.
├── data
│   ├── train
│   │   ├── videos
│   │   │   └── *.mp4
│   ├── public_test
│   │   ├── public
│   │   │   ├── videos
│   │   │   │   └── *.mp4
│   ├── public_test_2
│   │   ├── videos
│   │   │   └── *.mp4

To prepare image data for training and testing, run:

python3 utils/prep_data.py

Logging

This repository uses Weights & Biases for experiment tracking. See this for how to get started and log in with your account. Then, go to train.py and modify the wandb.init() line for your own project.

Training

To train with a specific model, run:

python3 train.py --backbone <any timm model>

Testing

To get predictions for the public test set 1, run

python3 infer.py --data data/public_test/public/images/

To get predictions for the public test set 2, run

python3 infer.py --data data/public_test_2/images/

Use the flag -h to see other training or inference arguments.