Refining Pseudo Labeling via Multi-Granularity Confidence Alignment for Unsupervised Cross Domain Object Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• State-of-the-art object detection methods struggle with generalization due to domain shift.
• Unsupervised cross-domain object detection aims to bridge this gap by transferring knowledge from labeled source domains to unlabeled target domains.
• Existing Mean Teacher methods show promise but are limited by suboptimal pseudo-labeling due to confidence misalignment.

Purpose of the Study:

• To address the challenge of confidence misalignment in pseudo-labeling for unsupervised cross-domain object detection.
• To propose a novel framework, Multi-Granularity Confidence Alignment Mean Teacher (MGCAMT), that simultaneously aligns confidence across category, instance, and image levels.
• To enhance the accuracy and generalization of object detectors in the presence of domain shift.

Main Methods:

• Developed Multi-Granularity Confidence Alignment Mean Teacher (MGCAMT) framework.
• Introduced Classification Confidence Alignment (CCA) using Evidential Deep Learning (EDL) to model category uncertainty and filter incorrect labels.
• Designed Task Confidence Alignment (TCA) to mitigate instance-level misalignment between classification and localization.
• Implemented Imagery Focusing Confidence Alignment (FCA) for balanced spatial layout perception without explicit label assignment.

Main Results:

• MGCAMT effectively alleviates confidence misalignment at category, instance, and image levels.
• The proposed CCA, TCA, and FCA components contribute to refined pseudo-labeling and improved teacher-student learning.
• Experimental results demonstrate significant performance improvements compared to existing state-of-the-art methods across various scenarios.
• The framework shows superior performance against large foundational models.

Conclusions:

• Confidence misalignment in pseudo-labeling is a critical bottleneck in unsupervised cross-domain object detection.
• MGCAMT offers a robust solution by integrating multi-granularity confidence alignment strategies.
• The proposed method significantly enhances object detection performance and generalization capabilities, outperforming current approaches.