Adaptive Bounding Box Uncertainty via Conformal Prediction
Published in ICCV - UNQCV Workshop, 2023
Recommended citation: Alexander Timans, Christoph-Nikolas Straehle, Kaspar Sakmann, Eric Nalisnick (2023). "Adaptive Bounding Box Uncertainty via Conformal Prediction." https://uncv2023.github.io/papers/ http://ksakmann.github.io/files/conformalbbox_uncv23.pdf
We quantify the uncertainty in multi-object bounding box predictions via conformal prediction. Using novel ensemble and quantile regression formulations, we are able to achieve per-class prediction intervals with guaranteed coverage that are adaptive to object size. We validate our approaches on real-world datasets (COCO, Cityscapes, BDD100k) for 2D bounding box localization, and achieve the desired coverage targets with sensibly tight intervals.