We study how machines can recognize, reconstruct, and act in complex visual environments. Our work spans open-ended vision-language recognition, 3D scene understanding, Gaussian-based scene reconstruction and relighting, and embodied perception for robotic and surgical systems.
Training-free recognition from internal vision–language evidence.
Given an image and region mask, Retrieve–Augment–Generate recognizes regions without predefined class lists or additional training, using internal evidence from frozen VLMs.
Understanding 3D environments beyond object categories.
3D scene understanding extends beyond assigning categories to objects by capturing richer semantics such as properties, affordances, and spatial relationships. This enables machines to better understand and interact with complex 3D environments, supporting applications such as robotics and autonomous systems.
Scene reconstruction and relighting with Gaussian splatting.
We study how images, depth estimation, normal estimation, and laser-scanned geometry can work together to reconstruct outdoor scenes and support changes in illumination.
Perception and multimodal reasoning for systems that must act in the physical world.
We study object recognition, tracking, spatial reasoning, and vision-language intelligence for robotic and surgical environments where perception must support real-time decisions and action.