Preview

PDF, English Download (13MB) | Terms of use

Abstract

While deep neural networks deliver state-of-the-art performance in object detection, their inherent tendency toward overconfidence compromises their reliability in safety-critical applications, necessitating robust methods for uncertainty quantification. Although full Bayesian inference would provide the most principled treatment of uncertainty, it is computationally impractical or even infeasiblefor modern largescale models and real-time detection pipelines. However, the application of Bayesian approximation techniques to complex, real-world object detection scenarios remains significantly underexplored, as existing literature focuses predominantly on simplified toy problems and lower-dimensional datasets. To address this gap, this thesis implements and evaluates Deep Ensembles and Monte Carlo Dropout within the state-of-the-art YOLOv8 architecture, assessing their ability to capture aleatoric and epistemic uncertainty across a corruption-augmented COCO dataset. Various Monte Carlo Dropout configurations with different dropout locations were explored; however, Deep Ensembles offer superior robustness and epistemic uncertainty estimation compared to Monte Carlo Dropout, which requires aggressive dropout in the detection head to remain effective.