DATR: Unsupervised Domain Adaptive Detection Transformer with Dataset-Level Adaptation and Prototypical Alignment.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Object detection models often struggle with domain shift, where performance degrades when applied to data from different distributions.
• Existing unsupervised domain adaptation methods for object detection primarily focus on aligning backbone or encoder features, neglecting the unique decoder structure of DETR-based models.
• Instance-level feature alignment in current methods is often class-agnostic, limiting improvements by overlooking category-specific distinctions.

Purpose of the Study:

• To develop a robust DETR-based object detector, named Domain Adaptive detection TRansformer (DATR), for unsupervised domain adaptation.
• To address the limitations of existing methods by introducing effective instance-level feature alignment within the decoder.
• To enhance the generalization ability of object detectors to new domains by considering dataset-level cues.

Main Methods:

• Proposed the Class-wise Prototypes Alignment (CPA) module to align cross-domain features in a class-aware manner.
• Introduced the Dataset-level Alignment Scheme (DAS) using contrastive learning to achieve global representation and improve inter-class feature distinguishability across domains.
• Incorporated a mean-teacher-based self-training framework to leverage pseudo-labels and mitigate domain bias.

Main Results:

• DATR demonstrated superior performance in unsupervised domain adaptation for object detection tasks.
• The proposed CPA module effectively aligns features class-wise, bridging the gap between object detection and domain adaptation.
• The DAS module enhanced global representation and inter-class feature distinctiveness, leading to improved generalization.

Conclusions:

• DATR significantly improves object detection performance under unsupervised domain adaptation by effectively aligning instance-level features within the decoder.
• The combination of class-wise and dataset-level alignment strategies, along with self-training, enhances model robustness and generalization.
• The developed method offers a promising direction for adapting object detectors to new, unseen domains without requiring labeled target data.