Paper accepted for publication in the proceedings of WACV2024

24.10.2023|15:38 Uhr

Our paper "Identifying Label Errors in Object Detection Datasets by Loss Inspection" by Marius Schubert, Tobias Riedlinger, Karsten Kahl, Daniel Kröll, Sebatsian Schönen, Siniša Šegvić and Matthias Rottmann has been accepted for publication in the proceedings of WACV2024!

In this work, we consider four types of bounding box label errors: missing labels, labels that do not correspond to any object, labels with wrong class, and labels with inaccurate localization. We show that they can be detected by instance-wise loss inspection. For benchmarking, we perturb labels from well-labeled datasets and show that we detect these errors, but we also show that we can find real labeling errors in common object recognition datasets using our method. The preprint is available at arxiv.org/pdf/2303.06999.pdf. In the conference version, we will add a theoretical justification for our method. Stay tuned and maybe meet us in Hawaii at WACV2024. This work was done in collaboration with TU Berlin, Control Expert, University of Zagreb / FER.