Exclusive style removal for cross domain novel class discovery

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Novel Class Discovery (NCD) is crucial for open-world learning, typically clustering unseen classes from unlabeled data using labeled data from the same domain.
• Existing NCD methods struggle when novel classes have different data distributions than labeled data, significantly compromising performance.

Purpose of the Study:

• To investigate and establish the solvability of NCD in cross-domain settings.
• To develop a method for improving NCD performance when labeled and unlabeled data distributions differ.

Main Methods:

• Introduced an exclusive style removal module to extract style information distinct from baseline features.
• The module acts as a plug-in, easily integrated with existing NCD methods.
• Established a fair benchmark for NCD research, considering backbone and pre-training strategy influences.

Main Results:

• The proposed style removal module effectively facilitates inference by removing style information.
• Demonstrated improved performance on novel classes with distributions different from the labeled set.
• Extensive experiments on three datasets validated the effectiveness of the style removal strategy.

Conclusions:

• Style information removal is a necessary condition for solving cross-domain NCD.
• The proposed plug-in module enhances NCD performance across different data distributions.
• The established benchmark provides a foundation for future cross-domain NCD research.