For long-term point tracking we make use of the objects_coordinates annotation in the MOVi-E dataset, along with ground-truth positions of objects, camera, and depth maps to generate point-tracks. Implementation-wise, create_point_tracking_dataset() is a simple wrapper around the default tfds Kubric loader which samples points and tracks them. For a demo, run python3 dataset.py. This demo requires installation of requirements_full.txt as it depends on tensorflow_graphics and matplotlib.