OpenOSR – General-Purpose Open-Source Tools for Open-Set Recognition

With the OpenOSR project, on one hand we would want to induce awareness that machine learning methods cannot be applied to any type of scenario, especially when unexpected inputs might need to be handled by such systems. On the other hand, we want to enable researchers at the University of Zurich (UZH) and beyond to make use of our methods for improving the classification under the presence of the unknown, to apply them on their own problem settings, and provide those as benchmarks to the research community.

When deployed in a wide range of applications, machine learning models need to handle unexpected inputs that they were not designed for. Open-Set Recognition (OSR) and Anomaly Detection (AD) are two research directions for developing methods that can consider and correctly treat such input. However, evaluation benchmarks for AD and OSR are usually either too simple or highly engineered, and limited to few input modalities, providing little insight into behavior of such algorithms in other modalities and more realistic applications. Our currently implemented toolbox OpenOSR has a focus on reproducibility of experiments and provides open-source implementations for several AD and OSR algorithms, as well as a few default small-scale and large-scale evaluation benchmarks that are mostly designed for image-based input modalities.

The goal of this project is to extend our OpenOSR framework to include more input modalities, implement machine learning tools including modern foundation models, test these on more realistic problem settings, and solve real-world problems as provided by collaboration partners.

If you have a machine learning application that requires to handle diverse and possibly unexpected inputs, and you want to be part of this exciting development, or you simply need help to evaluate our methods on your problem setting, please contact manuel.guenther@uzh.ch.

Project duration: 01.07.2025 - 31.06.2026

Contact: Prof. Dr. Manuel Günther

Project Lead

Since July 2020, Prof. Dr. Manuel Günther is Assistent Professor for Artificial Intelligence and Machine Learning at the Department of Informatics at the University of Zurich (UZH). There, he studies topics around the area of Deep Learning, which he often applies to images. Particular projects involve, for example, the identification of people based on their facial images, the development of methods to reduce bias in Deep Learning models, with the extension of image-based explainable AI, with the integration of insights of the visual system in the human brain into Deep Learning models, as well as with the recognition of the unknown.