The European Conference on Computer Vision (ECCV) is one of the world’s three leading conferences in the field of computer vision. The 19th edition took place in September 2026 in Malmö, Sweden.
The Regensburg Medical Image Computing (ReMIC) laboratory at OTH Regensburg was represented there with the paper ‘Distribution-Aware Feature Selection for Post-hoc Out-of-Distribution Detection’, authored by Max Gutbrod, a PhD student at ReMIC, together with co-authors David Rauber and Christoph Palm. Having a paper accepted at an A* conference – the highest category in the international ranking – for two years running is a remarkable achievement. The presentation in Malmö was made possible thanks to the support of the Bavarian Academic Forum for Health (BayWISS), which funded the travel expenses.
The work focuses on the detection of out-of-distribution (OOD) data – that is, image data that does not correspond to a model’s training data and which, in medical applications, can lead to clinically relevant errors. The key finding is that a small subset of carefully selected features already accounts for the majority of the discriminative power between in-distribution and out-of-distribution data. By focusing on these features, OOD detection becomes more reliable and, because less information needs to be processed, also faster. This applies to both medical image data and everyday photographs – meaning the method is effective beyond the field of medicine. The proposed pre-processing step can be widely applied to existing feature-based OOD methods.
In this way, OTH Regensburg is once again contributing to the discussion on the reliability of AI systems in safety-critical application areas such as medical image processing.
