Our research aims to advance systems capable of understanding and explaining the world, tackling tasks like object permanence and structural consistency using human-like context and cues. We would love to chat about this! Come see us at poster session 5 (Friday 11:00-14:00)! #NeurIPS 5/5

Building upon this benchmark, we also introduce a large-scale testing dataset of 10,000 scenes alongside a benchmark method. This is a collaborative work from the University of York between myself, @jadgardner.bsky.social and @willsmithvision.bsky.social 4/5

We introduce TAPVid-360; given query points as coordinates in the first frame, the goal is to track the 3D direction (in the camera coordinate frame) to the scene point. We aim for a model to predict in which relative direction a point is but, unlike 3D, not its distance. 3/5

Humans can create internal world models that complete the full sphere of surrounding information, even when only a fraction is currently visible. However, current AI vision systems lack this, often using an egocentric, frame-by-frame approach with poor memory for what is not immediately visible. 2/5