Vision systems that learn from as little labeled data as possible while staying honest about what they do not know — active learning and unsupervised accuracy estimation for classification, object detection, and semantic segmentation, including probabilistic multi-view 3D understanding.
Introduces a Beta distribution approximation for efficient uncertainty estimation in semantic segmentation, achieving significant annotation cost reductions through intelligent sample selection.
A self-supervised approach to 3D geometric data analysis that removes the need for manual annotation in mesh segmentation tasks.
A representation and active learning framework built for heavily imbalanced datasets, validated on COVID-19 chest X-ray classification.
Clinical validation showing a substantial reduction in radiologist annotation burden while maintaining diagnostic accuracy. Received the Best Presentation Award at FICC 2022.
Active learning frameworks spanning classification, object detection, and semantic segmentation, deployed to cut labeling budgets on industrial and medical datasets.
Methods for estimating how well a deployed vision model is performing without access to labeled evaluation data — essential when distributions shift after deployment.
A probabilistic active learning framework for 3D scene understanding that fuses multiple viewpoints to improve segmentation accuracy while quantifying prediction uncertainty.
Vision systems that provide reliable confidence estimates for high-stakes applications, letting clinicians know which predictions warrant review.
Industrial vision systems for manufacturing quality control, including real-time defect recognition and image processing for 3D CT composite data.