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.
Loss functions and representations that make neural networks respect the connectivity and topology of what they predict.
Standard pixel-wise losses ignore whether a predicted structure is actually connected. We design topology-aware objectives — built on persistent homology, region separation and path enforcement — that teach networks to preserve the global structure of their outputs, not just local accuracy.
Detecting thin, connected networks — roads, blood vessels, neurons, cracks — where a single missed pixel breaks the whole graph.
Linear structures form connected networks whose usefulness depends on correct connectivity. We develop methods that delineate these structures in 2D and 3D while explicitly enforcing their connectivity, with applications from aerial road mapping to neuron tracing and structural crack detection.
Models that know when they might be wrong — calibrated, efficient uncertainty for high-stakes decisions.
Deployment in medicine and safety-critical settings demands models that quantify their own confidence. We study how to obtain reliable, efficient uncertainty estimates, including methods that enable uncertainty in iterative neural networks without expensive ensembles.
Bringing structure-aware and uncertainty-aware deep learning to clinical imaging.
We apply our methods to medical imaging problems — from delineating anatomical structures to detecting disease signatures such as bronchiolitis obliterans syndrome in chest CT — emphasizing connectivity, reliability and clinical trust.
Part-based implicit neural representations for composite 3D shapes and optimization.
We investigate neural representations of 3D geometry, including part-based implicit fields that parametrize composite shapes and support shape optimization, bridging computer vision and geometry processing.
Forecasting agent motion in autonomous-driving scenes by modeling interactions, uncertainty and multimodal future behavior.
Autonomous systems must predict not only where agents may move, but how their futures depend on one another. We study interaction-aware motion forecasting methods that model social and traffic interactions, multimodal futures, uncertainty and precedence structure, aiming to improve reliable prediction in complex autonomous-driving scenes.
Methods & ethos
Across every project we ask the same two questions: does the model capture the structure of the problem — the connectivity, the geometry, the topology — and does it know when it might be wrong? We favour methods that are principled, reproducible and honest about their limits, because the applications we care about leave little room for confident mistakes.
