This repository builds a simulator that tracks the gaze of all the people in front of a kinect camera using output from a Gaze Detection model and position of people from kinect.
pykinect_azure pygame opencv-python numpy matplotlib
clone the repository with
git clone https://github.com/eliird/GazeDetectionSim python run.py
The details of implementation can be found in the Simulator class in simulator.py
The image below shows the architecture of the gaze detection model 
The model is based on the gaze 350 but we replaced the backbone with the convNext models. We replace the backbone and retrain the model. The code for the backbone can be found in /convnext directory. And the architecture of the complete model can be found in model.py file.
The retrained model has the Accuracy of 11.6 degrees on test dataset provided in gaze360 dataset.
Test involved asking people to walk in front of the camera at distance of 1m, 1.5m and 2m while looking at the either one of the three objects present in fron tof them for 1 minute. They were randomly asked to look at different object every 5 seconds.
Refere the image below
Total of 3 subjects were each asked to walk at all three distances and the average of all 3 is recorded below
Accuracy for each distance TODO Update the accuracies
1.0m --- 2.12 %
1.5m --- 2.12 %
2.0m --- 2.12 %
The weights of the model can be downloaded with gdown
gdown --id 1pv_kSHkVqar8E3WEqVCZAdoPlFmusHZl on linux
or use this link to download the weights of the trained model.