Biased Class disagreement: detection of out of distribution instances by using differently biased semantic segmentation models

Autonomous driving heavily relies on accurate understanding of the surrounding environment, which is facilitated by semantic segmentation models that classify each pixel in an image. However, training these computer vision models using available datasets often fails to capture the diverse conditions and objects that can be encountered during a trip. Adverse weather conditions and the presence of Out-of-Distribution (OOD) instances, such as wild animals and debris, are common challenges in autonomous driving. Unfortunately, current models struggle to perform well in unseen conditions. To address these limitations, this paper proposes a comprehensive approach that integrates uncertainty quantification and bias reinforcing within the framework of Unsupervised Domain Adaptation (UDA). Our approach leverages multiple models with diverse biases, aiming to assign high-confidence predictions to OOD instances by mapping them to the selected prior semantic category. Extensive evaluations on the MUAD dataset demonstrate the effectiveness of our approach in improving performance and robustness against OOD instances. Notably, our approach achieves outstanding results, securing the first position in the MUAD challenge.