The practical deployment of deep learning-based object detectors, despite their benchmark successes, remains vulnerable to performance degradation under domain shifts—mismatches between the training and testing data distributions. These shifts, stem...
The practical deployment of deep learning-based object detectors, despite their benchmark successes, remains vulnerable to performance degradation under domain shifts—mismatches between the training and testing data distributions. These shifts, stemming from real-world variations in lighting, weather, scene composition, and sensor characteristics, call for robust adaptation mechanisms to ensure reliable operation. This dissertation addresses these challenges through a comprehensive study of adaptation strategies, spanning from offline preparation to dynamic online adaptation at inference time.
We begin by revisiting Unsupervised Domain Adaptation (UDA), which enables learning in target domains where labels are unavailable. Object detection presents unique challenges as it requires not only classification but also accurate localization, which involves regressing localization-specific continuous values (e.g., bounding box offsets). Most existing UDA methods focus on category-level alignment, which—while generally effective for classification—may fall short in preserving the spatial precision required for detection. To overcome this, we propose Offset-Aware Domain Adaptation (OADA), a novel offline approach that conditions feature alignment directly on offset predictions. By explicitly accounting for localization-sensitive variations, OADA significantly improves the effectiveness of feature alignment and strengthens the generalization capacity of detectors across domains.
Recognizing that offline strategies alone are insufficient in dynamically changing environments, we further explore Continual Test-Time Adaptation (CTA), where models must incrementally adapt to unlabeled test streams without access to source data. Existing online adaptation methods are typically designed for classification and fail to account for the structural and architectural characteristics of object detection. In particular, reliance on Batch Normalization updates proves unstable under small batch sizes and incompatible with modern detectors based on LayerNorm or Transformer backbones.
To address this, we introduce a lightweight, architecture-agnostic CTA framework tailored to object detection. Our method freezes the pre-trained backbone and updates only compact adaptor modules, ensuring efficiency and stability while preserving source-domain knowledge. We further incorporate class-wise alignment and a selective update mechanism that dynamically determines when adaptation is necessary, enabling the model to respond effectively to evolving domains without incurring excessive computational cost.
In summary, this dissertation presents a unified and detection-aware framework for domain adaptation under both offline and online settings. By combining offset-aware alignment with a principled continual adaptation strategy, we significantly enhance the robustness and deployability of object detectors in complex, real-world environments characterized by continuous domain shifts.