article-journal

A post-training framework that aligns DINO features across RGB, depth, segmentation, and other visual modalities while preserving discriminative semantics.

A resource-efficient extension of SAM2 that grounds open-ended language prompts into precise segmentation masks.

A generalist forecasting framework that predicts future frozen visual features and decodes them for multiple downstream tasks.

A study showing that large-scale self-supervised video models improve consistently on spatial and temporal 4D vision tasks as model size grows.

An open vision-language model designed as a versatile base model for transfer across a wide variety of open-world tasks.

An unsupervised module for aligning object representations across time so entities can be tracked between segmented frames.

An unsupervised model that jointly learns attention masks and region reconstructions to decompose scenes into meaningful objects and background elements.