Automated labeling pipeline for 2D object detection

Object detection remains limited by its reliance on large-scale, high-quality labeled datasets, an assumption rarely met in real-world applications. Practical deployments often involve noisy or sparse labels, long-tailed class distributions, and shifting environmental conditions that degrade performance. This thesis introduces a data-efficient learning pipeline that combines uncertainty estimation, failure recognition, active learning (AL), and semi-supervised learning (SSL) to systematically reduce the need for manual labeling and improve model robustness under imperfect data conditions. Uncertainty estimation enables object detectors to quantify the reliability of their predictions, providing a foundation for risk-aware decision-making. Building on this, uncertainty-based failure recognition allows the system to automatically detect and filter out incorrect predictions, improving safety and enabling reliable auto-labeling without human supervision. AL identifies the most informative samples for labeling, often using model uncertainty as a proxy, thereby focusing labeling effort where it yields the greatest performance improvement. With AL minimizing the need for manual labeling and failure recognition identifying trustworthy predictions, SSL can effectively leverage large volumes of unlabeled data in a self-training loop, where the model refines itself using its own confident predictions. The thesis makes four primary contributions: (1) introducing methods to rapidly and accurately propagate and calibrate aleatoric localization uncertainty; (2) proposing an uncertainty-driven thresholding framework enabling automatic recognition of detection failures and auto-labeled images; (3) presenting a metric that supports the development of AL methods, increasing their effectiveness and promoting the practical real-world usability of AL with a case study using uncertainty-based AL methods; and (4) developing data-centric methods within SSL frameworks to address class imbalance, erroneous labels, and confirmation bias from pseudo-labeling. Collectively, these contributions improve the efficiency, reliability, and scalability of object detectors while substantially reducing the reliance on high-quality manual labels. The effectiveness of the proposed approaches is validated through experiments on real-world autonomous driving datasets.