fd_shifts

This repository contains a framework for benchmarking failure detection methods under distribution shifts in image classification featuring implementations of baseline methods, data sets and metrics. Additionally to a wide range of built-in methods, this framework can be used to evaluate and compare new methods, models and datasets.

Detecting failures should be seen in the context of the overarching goal of preventing silent failures of a classifier by means of confidence scoring functions. © Paul F. Jäger

Find fd_shifts on Helmholtz Imaging Connect
Authors: Paul F. Jäger, Carsten T. Lüth, Lukas Klein, Till J. Bungert