Instance-Invariant Domain Adaptive Object Detection Via Progressive Disentanglement.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection models struggle with generalization when training and testing data originate from different domains.
• Existing domain adaptation methods often overlook the influence of domain-specific information within aligned features.

Purpose of the Study:

• To develop a method for extracting instance-level, domain-invariant features for improved cross-domain object detection.
• To disentangle domain-invariant features from domain-specific features to enhance generalization.

Main Methods:

• A progressive disentangled mechanism, comprising base and progressive disentangled layers, was proposed to decompose features.
• Instance-invariant features were extracted using the Region Proposal Network (RPN) on the output of the progressive disentangled layer.
• A detached optimization strategy was employed for end-to-end model training.

Main Results:

• The proposed method achieved performance improvements of 2.3%, 3.6%, 4.0%, and 2.0% over the baseline across four domain-shift scenarios.
• Visualization analysis confirmed the model's effective disentanglement of domain-invariant and domain-specific features.

Conclusions:

• The novel feature disentanglement approach significantly enhances object detection performance in cross-domain settings.
• The method demonstrates robust generalization capabilities by effectively separating domain-specific information.