Domain Shift Preservation for Zero-Shot Domain Adaptation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Learning-based image processing models struggle with domain shift due to differing data distributions.
• Domain adaptation techniques aim to transfer knowledge from source to target domains.
• Zero-shot domain adaptation (ZSDA) is challenging as target domain data is unavailable during training.

Purpose of the Study:

• To propose a novel method for zero-shot domain adaptation (ZSDA) that effectively transfers knowledge to unseen target domains.
• To address the limitations of current ZSDA methods by leveraging domain shift preservation across tasks.

Main Methods:

• A strategy of domain shift preservation across tasks is employed.
• Domain shift is learned from an irrelevant task with available data from both source and target domains.
• Two coupled generative adversarial networks (CoGANs) capture joint dual-domain distributions.
• A generative adversarial network (GAN) explicitly models the domain shift.

Main Results:

• The proposed method successfully learns models for unseen target domains.
• Experimental results demonstrate satisfactory performance in image classification and semantic segmentation.
• The method effectively transfers various types of domain shifts across different tasks.

Conclusions:

• The domain shift preservation strategy is effective for zero-shot domain adaptation.
• The proposed CoGANs and GAN-based approach provides a robust solution for cross-domain knowledge transfer.
• This work advances the capabilities of machine learning models in handling diverse data distributions without target-specific training data.