paper-conference

An information-theoretic framework for Bayesian active learning and optimization in censored-data regimes.

A general-purpose perception model built from a pretrained video diffusion backbone and steered across vision tasks with text instructions.

A modular study of image-to-3D pipelines showing that camera-relative object coordinates provide an effective intermediate shape representation.

An empirical comparison of image- and video-diffusion representations across classification, action recognition, depth estimation, and tracking.

A self-supervised video model whose latent tokens move off the image grid to represent and track scene elements consistently through time.

A method for controlling the 3D pose and placement of individual objects in image diffusion models using per-object Neural Assets.

We improve pretrained captioning and classification of 3D objects via visually grounded aggregation of VLM responses.

A video scene model which separates the time-invariant, object-level contents of the scene from global time-varying elements such as viewpoint.

An empirical study showing that a standard model-free reinforcement-learning agent can learn behaviors commonly associated with planning.

An unsupervised iterative-inference method that jointly segments scenes into objects and learns disentangled object representations.