Mean teacher based on class prototype contrast for domain adaptive object detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptive object detection (UDAOD) applies models trained on labeled source data to unlabeled target data.
• The mean teacher framework is common in UDAOD but struggles with false positives and incomplete pseudo-labels due to domain differences.

Purpose of the Study:

• To propose a novel student-teacher framework, Prototype Contrast Mean Teacher (PCMT), to enhance UDAOD performance.
• To address challenges of feature discrepancies and insufficient pseudo-labels in cross-domain object detection.

Main Methods:

• PCMT utilizes class prototypes for contrastive learning to preserve intra-class features and align features across domains.
• A pseudo-label filtering method based on bounding box localization is introduced to improve label quality.

Main Results:

• PCMT demonstrates superior performance in UDAOD across various domain adaptive conditions.
• On the Cityscapes → BDD100K dataset, PCMT achieved a 43.5% mAP, outperforming the state-of-the-art by 5.0%.

Conclusions:

• The proposed PCMT framework effectively improves unsupervised domain adaptive object detection.
• PCMT's approach to class prototypes and pseudo-labeling offers a significant advancement in cross-domain computer vision tasks.