by acknowledging to the NvGaze license you can use the provided 3D eye model to render a customized synthetic data set from a virtual camera position that fits your application. As described in the paper changing the camera position has significant impact on the appearance of the eye and the resulting accuracy of pupil localization. Please be aware that the 3D model includes the eye but no face mesh due to the license of the face scans. Therefore, as you pointed out we provide our complete data set to work with (including synthetic data and real captured data).
Our trained model is only available with a business license.
If you want to train your own network we provide accelerated inference via TensorRT that comes with the Jetpack.