Awesome

CUDA-GHR: Controllable Unsupervised Domain Adaptation for Gaze and Head Redirection

• Published at WACV2023!
• Authors: Swati Jindal, Xin Eric Wang
• Link to paper

Requirements

We used Python 3.7.10 and torch 1.18.1 to test our experiments. We ran our codebase on Ubuntu 20.04.

To install all the packages:

pip install -r requirements.txt

Usage

Data

Download the three datasets: GazeCapture, MPIIFaceGaze, Columbia.

To pre-process the datasets, please use this repository and follow instructions provided to generate eye-strip images for FAZE. Put the h5 files in the data folder.

Train

Create a config json file similar to configs/config_gc_to_mpii.json describing all the training parameters and paths to the input files.

To train the task network, run this command:

python train_tasknet.py --config-json configs/config_tasknet.json

To train and evaluate the CUDA-GHR model in the paper, run this command:

GazeCapture → MPIIGaze:

python train_cudaghr.py --config_json configs/config_gc_to_mpii.json

GazeCapture → Columbia:

python train_cudaghr.py --config_json configs/config_gc_to_col.json --columbia

The training images, losses and evaluation metrics will be loggged in Tensorboard. We also save generated images in the save folder.

Evaluate

To evaluate CUDA-GHR model, run this command:

python eval_cudaghr.py --model_path <path to model> --config_json <path to config file> --test_people <subset to test>

Add '--columbia' option to test on Columbia dataset.

Pre-trained Models

You can download pretrained models here:

• GazeCapture → MPIIGaze
• GazeCapture → Columbia
• ResNet-50 task model

Acknowledgement

The code is adapted from FAZE and STED-Gaze. We thank authors for their awesome work!!