Automatically Discovering Novel Visual Categories With Adaptive Prototype Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Novel Category Discovery (NCD) addresses the challenge of identifying unknown classes in datasets with pre-existing known classes.
• Real-world scenarios often present partial data, making NCD crucial for comprehensive understanding.
• Existing NCD methods struggle with incomplete annotations for novel categories.

Purpose of the Study:

• To propose a novel adaptive prototype learning method for effective Novel Category Discovery (NCD).
• To emphasize category discrimination and mitigate issues arising from missing annotations in novel classes.
• To develop a robust feature extractor capable of handling both known and unknown image categories.

Main Methods:

• Prototypical representation learning using a robust feature extractor enhanced by self-supervised learning and adaptive prototypes.
• Prototypical self-training stage to refine pseudo-labels and train a final classifier for category clustering.
• Leveraging adaptive prototypes to improve instance and category discrimination.

Main Results:

• The proposed method demonstrates state-of-the-art performance across four benchmark datasets.
• Achieved high effectiveness and robustness in novel category discovery tasks.
• Successfully handled images from both base and novel categories with improved discrimination.

Conclusions:

• Adaptive prototype learning offers a powerful approach to Novel Category Discovery (NCD).
• The method effectively addresses the challenge of missing annotations in real-world datasets.
• The developed feature extractor and training strategy significantly advance the state-of-the-art in NCD.