Scale-Consistent and Temporally Ensembled Unsupervised Domain Adaptation for Object Detection

Unsupervised Domain Adaptation for Object Detection (UDA-OD) aims to adapt a model trained on a labeled source domain to an unlabeled target domain, addressing challenges posed by domain shifts. However, existing methods often face significant challenges, particularly in detecting small objects and over-relying on classification confidence for pseudo-label selection, which often leads to inaccurate bounding box localization. To address these issues, we propose a novel UDA-OD framework that leverages scale consistency (SC) and Temporal Ensemble Pseudo-Label Selection (TEPLS) to enhance cross-domain robustness and detection performance. Specifically, we introduce Cross-Scale Prediction Consistency (CSPC) to enforce consistent detection across multiple resolutions, improving detection robustness for objects of varying scales. Additionally, we integrate Intra-Class Feature Consistency (ICFC), which employs contrastive learning to align feature representations within each class, further enhancing adaptation. To ensure high-quality pseudo-labels, TEPLS combines temporal localization stability with classification confidence, mitigating the impact of noisy predictions and improving both classification and localization accuracy. Extensive experiments on challenging benchmarks, including Cityscapes to Foggy Cityscapes, Sim10k to Cityscapes, and Virtual Mine to Actual Mine, demonstrate that our method achieves state-of-the-art performance, with notable improvements in small object detection and overall cross-domain robustness. These results highlight the effectiveness of our framework in addressing key limitations of existing UDA-OD approaches.