Visual AI for Extended Reality

Self-supervised learning

The trend goes towards massive datasets with billions of images, an amount that cannot be annotated manually. To this end, we develop self-supervised algorithms that recover structure automatically, such as masks of the object boundaries, keypoint locations, or the skeleton that relates parts of an object as a graph.
A recent focus is moving from low-level geometric properties to high-level descriptions. For instance, to model the behavior of animals and persons and to relate their actions and interactions with the environment. To this end, we leverage large generative models and give them additional control beyond text input.