Evaluating and Advancing Object Detection Models for Robust Generalization under Domain Shifts

Abstract

Modern object detectors are usually evaluated using accuracy metrics such as Average Precision, which provide limited insight into the reliability and calibration of predictions, particularly when models are deployed under distribution shift. This thesis explores uncertainty-aware object detection and its behavior under cross-dataset transfer in autonomous driving scenarios. The first part of it establishes a controlled evaluation framework following the CertainNet protocol. Gaussian YOLOv3, CenterNet, and CertainNet are trained on the KITTI dataset and evaluated both in-domain and under transfer to BDD100K and nuImages. The results show that structured uncertainty modeling improves prediction calibration and spatial uncertainty estimation while also maintaining competitive detection accuracy. Building on this baseline, the second part of the thesis explores possible extensions within a semi-supervised teacher–student detection framework. Internal uncertainty supervision, uncertainty-guided pseudo-label filtering, and representation alignment using DINOv2 features are investigated through a systematic ablation study. The experiments that were conducted show that representation alignment provides the most significant improvements, especially in cross-domain detection performance, while maintaining reliable uncertainty estimates.