Foregroundness-Aware Task Disentanglement and Self-Paced Curriculum Learning for Domain Adaptive Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptive object detection (UDA-OD) aims to detect objects across different domains without labeled target data.
• Existing UDA-OD methods often integrate adaptation modules directly into detectors, potentially degrading performance.
• A key challenge is balancing generalization ability with maintaining detection accuracy.

Purpose of the Study:

• To propose an effective framework, FA-TDCA, to address the limitations of current UDA-OD methods.
• To disentangle the UDA-OD task into independent subtasks for improved knowledge transfer and performance.
• To introduce a novel 'foregroundness' metric for evaluating location confidence and improving pseudo-label quality.

Main Methods:

• Developed the foregroundness-aware task disentanglement and self-paced curriculum adaptation (FA-TDCA) framework.
• Disentangled UDA-OD into source detector pretraining, classification adaptation, location adaptation, and target detector training.
• Introduced a 'foregroundness' metric to assess location confidence and combined it with classification confidence for label assessment.
• Employed a self-paced curriculum learning paradigm to gradually enhance pseudo-label quality for target samples.

Main Results:

• The proposed FA-TDCA framework effectively disentangles UDA-OD tasks, enabling efficient knowledge transfer.
• The novel foregroundness metric aids in assessing and improving the quality of object detection proposals.
• Self-paced curriculum adaptation progressively refines pseudo-labels, boosting adaptation performance.
• Achieved state-of-the-art results on four cross-domain object detection benchmarks.

Conclusions:

• FA-TDCA offers an effective approach to unsupervised domain adaptive object detection by disentangling tasks.
• The foregroundness metric and curriculum learning strategy contribute to improved adaptation and detection performance.
• This framework successfully enhances generalization ability while preserving high detection accuracy across domains.