We build computer-vision and deep-learning methods that respect the geometry, topology and uncertainty of the visual world — and put them to work in medicine, remote sensing and beyond.
We develop computer vision methods that explicitly account for topological properties such as connectivity, continuity, and the preservation of meaningful structures. Our work focuses on topology-aware learning for curvilinear and network-like structures in both 2D and 3D, with applications including vessels, neuronal structures, road networks, and other connected visual patterns.
We develop methods for estimating predictive uncertainty and improving the reliability and trustworthiness of deep learning models. Our research focuses on calibrated confidence, uncertainty-aware prediction, robustness, and identifying unreliable outputs, particularly in settings where dependable model behavior is critical.
We study neural representations for modeling, understanding, and manipulating complex 3D shapes. Our work includes implicit representations and part-based approaches for shape reconstruction, generation, parametrization, editing, and optimization, with an emphasis on representing geometry in flexible and controllable ways.
We develop methods for predicting the future trajectories and motion of agents in dynamic environments. Our work spans human trajectory forecasting and autonomous-driving scenarios, with a focus on multimodal prediction, interactions between agents, and modeling plausible future behavior in complex scenes.
