Unsupervised Domain Adaptive Object Detection via Semantic Consistency and Compactness Learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised domain adaptive object detection aims to improve model robustness in new domains without labeled data.
• Existing methods struggle with holistic feature consistency and reliable category feature compactness.
• Challenges include inefficient style matching, semantic discrepancy, poor sample quality, and noisy contrastive learning.

Purpose of the Study:

• To propose a novel Semantic Consistency and Compactness Learning (SCCL) network.
• To address the limitations of insufficient/inefficient consistency learning and unreliable compactness learning in unsupervised domain adaptation.
• To enhance feature transferability and discriminability for robust object detection.

Main Methods:

• Introduced a Visual Adaptation-guided Semantic Alignment (VSA) module for efficient feature consistency learning via feature adaptation and adversarial-free self-supervised feature disentanglement.
• Developed a plug-and-play Instance Center-Contrastive (ICC) head to address unreliable compactness learning by enhancing pseudo-label quality, improving sample storage/updating, and refining the contrast paradigm.
• Leveraged the mutual reinforcement between VSA and ICC.

Main Results:

• The proposed SCCL network demonstrated superior adaptability and robustness in unsupervised domain adaptive object detection.
• Achieved significant improvements across four benchmark datasets.
• The VSA and ICC modules effectively enhanced feature transferability and discriminability.

Conclusions:

• The SCCL network effectively overcomes key challenges in unsupervised domain adaptive object detection.
• The VSA and ICC modules provide a robust framework for learning feature consistency and compactness.
• SCCL offers a promising approach for enhancing model robustness in diverse target domains.