This code is for the paper 360-Degree Gaze Estimation in the Wild Using Multiple Zoom Scales. By using this code you agree to terms of the LICENSE.

Use the conda to create a new environment using the given .yml file. In case conda is not installed on your system, you can install it from here. Once conda is installed, use the following code to create an environment with all dependencies installed.

conda env create -f multizoomgaze_env.yml

Download the checkpoint files from here.

Use this notebook to predict gaze direction in a random image.

Register here which will then give you access to the database.

MSA+Seq

python run.py --model_type=NonLstmSinCosModel  --enable_time --checkpoints_path=CKECKPOINT_DIRECTORY/ --source_path=/data/GAZE360/imgs/ --evaluate

MSA

python run.py --model_type=NonLstmSinCosModel --checkpoints_path=CKECKPOINT_DIRECTORY/ --source_path=/data/GAZE360/imgs/ --evaluate

MSA+raw

python run.py --model_type=NonLstmMultiCropModel --checkpoints_path=CKECKPOINT_DIRECTORY/ --source_path=/data/GAZE360/imgs/ --evaluate

Pinball Static

python run.py --model_type=StaticModel --checkpoints_path=/home/ashesh/gaze_final_checkpoints/ --evaluate

Here, checkpoints_path is the directory where you've saved the trained checkpoint files. source_path is the directory which contains gaze360 data.

Just remove the --evaluate token from the command for evaluating the model performance which is given in the previous section. In this case, checkpoints_path will be the path where your checkpoints will get saved.