Publications

A hybrid long-video reasoning framework that proposes action-grounded hypotheses efficiently, then verifies only sparse RGB evidence with an expensive VLM.

A unified egocentric encoder trained via hierarchical distillation across ego-exo viewpoints, modalities, and foundation models, using representation-specific proxies as mediators.

The first Mamba-based architecture for temporal action detection in long untrimmed videos — only 17M parameters, trainable on an NVIDIA Jetson Nano.

A large language-vision model that incorporates 3D poses and object trajectories to understand the spatiotemporal relationships within daily living activities.

Introduces 3D skeletons into the vision-language embedding space to enable effective zero-shot learning for activities of daily living.

3D latent-controlled diffusion for identity-preserving face swapping.

A study of where video understanding stands with vision-language foundation models, and where it should go next.