Cascading Alignment for Unsupervised Domain-Adaptive DETR with Improved DeNoising Anchor Boxes

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Transformer-based object detection models, such as DINO, achieve state-of-the-art results.
• These models face challenges in new scenarios with differing imaging conditions and no annotated data (domain shift).

Purpose of the Study:

• To propose an unsupervised domain adaptive method for DINO to address the domain shift problem.
• To enable effective object detection in novel environments without requiring target domain annotations.

Main Methods:

• Introduced unsupervised domain adaptive DINO via cascading alignment (CA-DINO).
• Employed attention-enhanced double discriminators (AEDD) for local-global context alignment and domain discrepancy reduction.
• Utilized weak-restraints on category-level token (WROT) to minimize second-order statistics differences between domains.

Main Results:

• CA-DINO demonstrated effectiveness in unsupervised domain adaptation.
• Achieved a significant 41% relative improvement over the baseline on the Foggy Cityscapes benchmark.
• The approach is trained end-to-end, showing robust performance across challenging benchmarks.

Conclusions:

• CA-DINO successfully alleviates the domain shift problem in transformer-based object detection.
• The proposed method offers a viable solution for deploying object detection models in unseen environments.
• This work advances unsupervised domain adaptation techniques for advanced computer vision tasks.