Sebastien Gerchinovitz - Conformal prediction for object detection

Conformal prediction is a family of simple statistical methods to evaluate the uncertainty of black-box prediction models, with guaranteed risk bounds. A common limitation of such methods is that the uncertainty is modelled through a one-dimensional parameter. In this work, we address the problem of constructing reliable uncertainty estimates for object detection, in which the learning tasks and the associated predictive uncertainties are more complex. We first reduce the setting to a sequential two-task learning problem with two uncertainty parameters. Then we propose an extension of the CRC algorithm and derive associated risk bounds. Finally, we show how to instantiate this algorithm with classical deep object detectors, and illustrate the method on the COCO dataset. This is a joint work with Léo Andéol, Luca Mossina, and Adrien Mazoyer.