Self-Adversarial Disentangling for Specific Domain Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain adaptation methods often struggle with specific dimensional shifts (e.g., fog, rainfall).
• Existing approaches lack explicit prior knowledge of domain shifts, limiting adaptation performance.
• The intra-domain gap, caused by varying domainness, is a critical challenge in specific domain adaptation.

Purpose of the Study:

• To introduce and address the problem of Specific Domain Adaptation (SDA).
• To propose a novel framework, Self-Adversarial Disentangling (SAD), to mitigate the intra-domain gap.
• To improve adaptation performance by disentangling domainness-specific and domainness-invariant features.

Main Methods:

• Proposed the Self-Adversarial Disentangling (SAD) framework for Specific Domain Adaptation (SDA).
• Introduced a domainness creator to enrich the source domain with supervisory signals.
• Designed a self-adversarial regularizer and two loss functions to disentangle latent representations.

Main Results:

• Successfully disentangled latent representations into domainness-specific and domainness-invariant features.
• Mitigated the intra-domain gap effectively within the Specific Domain Adaptation (SDA) setting.
• Achieved consistent improvements over state-of-the-art methods in object detection and semantic segmentation.

Conclusions:

• The SAD framework offers a plug-and-play solution for Specific Domain Adaptation (SDA) without inference-time costs.
• The proposed method effectively addresses challenges posed by specific dimensional domain shifts.
• SAD demonstrates significant performance gains in computer vision tasks like object detection and semantic segmentation.