Visual intelligence that connects perception, reasoning, and action.
Our group develops visual intelligence for understanding people, events, and dynamic environments. General video understanding is a central theme: learning how appearance, motion, geometry, and language jointly convey what happens over time. We study vision-language models and multimodal reasoning, with an emphasis on temporal grounding, long-video understanding, and whether model predictions are supported by visual evidence.
Complementing this work, we study human pose and mesh recovery, motion modeling, and 3D/4D scene representations. These methods capture the structure and dynamics of people and their surroundings. Our research in generative modeling, neural rendering, and Gaussian splatting connects such representations with reconstruction, synthesis, and editing, bringing visual understanding and generation closer together.
Building on these foundations, we study predictive world models and physical and agentic AI to connect perception and reasoning with action. We investigate how models can anticipate outcomes, plan, use tools, and learn from execution feedback. This work also addresses the systems around models, including context management, memory, and tool interfaces, with an emphasis on evaluating reliability across tasks and environments.
Efficient learning and inference, together with federated and privacy-preserving methods, address the computational and data-sharing constraints that shape practical AI systems. We also pursue distinct research questions in medical imaging and healthcare, geospatial AI and remote sensing, image reconstruction, and foundational learning. These areas both benefit from and inform our broader work in representation learning and multimodal intelligence.
Understanding visual evidence, representing people and scenes, and generating visual content.
We study vision-language models and multimodal learning across images, video, text, and domain-specific data, including grounding and evaluation of reasoning fidelity.
We study long-form video understanding, temporal grounding, motion reasoning, and anomaly detection, supported by datasets and benchmarks.
We study 3D human pose and mesh recovery, motion modeling and generation, and human behavior analysis, including robustness across viewpoints and domains.
We study controllable generation and editing of images, video, and 3D content, including diffusion models, alignment, and geometry-aware synthesis.
Learning under computational, communication, and data-sharing constraints.
We study efficient model adaptation, compression, and dynamic computation to reduce the resources needed for vision and foundation models.
We study federated and privacy-preserving learning, including personalization, multimodal adaptation, and training across heterogeneous data and devices.
Distinct scientific questions in healthcare and geospatial intelligence, informed by shared learning methods.
We study biomedical foundation models, medical image analysis, computational pathology, and video or motion analysis for health assessment.
We study cross-view geolocalization, remote sensing, and geospatial foundation models using ground-level, aerial, and satellite imagery.