This project is for video-based gaze estimation of instantaneous frames network.
The code is built with following libraries:
- PyTorch 1.0 or higher
- TensorboardX
- tqdm
- scikit-learn
- pandas
- Python 3.x
To train the model(s) in the paper, run this command:
python main.py <train_video_path> \
--arch <resnet-backbone> \
--lr 0.002 --lr_steps 40 70 --epochs 100 \
--batch-size 128 --dropout 0.3 --consensus_type=avg --npb
The results of the training are saved as a .pth file format.
To test the trained models from video-based gaze dataset, you can run:
python test.py <test_video_path> \
--weights= <pth_path>\
--batch_size=32
W. Liu, [email protected]