The implementation of Arxiv paper for gaze estimation task.
Run the commands below from this repository's reconstruction/gaze/ directory; all input and output paths are relative to it.
-
Download the dataset and put it under
data/ -
Download eyes3d.pkl and put it under
assets/ -
Download pretrained checkpoint and put it under
assets/ -
Install libraries:
pip install timm pytorch-lightning==1.8.1 albumentations==1.3.0
After downloading the pre-trained checkpoint above,
python test_gaze.py assets/latest_a.ckpt
test_gaze.py reads images from assets/images/ and writes side-by-side input and gaze visualizations to outputs/, using the original filenames. Create assets/images/ and add your input images before running it.
python trainer_gaze.py
trainer_gaze.py reads train.rec, train.idx, val.rec, and val.idx from data/gaze_refine/ by default (override with --root). It saves checkpoints to work_dirs/gaze/ and TensorBoard logs under work_dirs/gaze/logs/.
