Virtual Category Learning: A Semi-Supervised Learning Method for Dense Prediction With Extremely Limited Labels.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised learning (SSL) is crucial due to the high cost of labeled data.
• Pseudo-labeling in SSL struggles with confusing samples, risking model generalization or confirmation bias.

Purpose of the Study:

• To propose a novel method for effectively utilizing confusing samples in semi-supervised learning.
• To improve model generalization and embedding space quality without label correction.

Main Methods:

• Introduced Virtual Category (VC) assignment for confusing samples, enabling safe optimization without concrete labels.
• VCs provide an upper bound for inter-class information sharing, enhancing the embedding space.
• Evaluated the method on semantic segmentation and object detection tasks.

Main Results:

• The proposed VC learning significantly surpasses state-of-the-art performance.
• Performance gains are particularly notable when limited labeled data is available.
• Demonstrated effectiveness in mainstream dense prediction tasks.

Conclusions:

• Virtual Category learning offers a proactive and effective approach to handling confusing samples in SSL.
• VC learning enhances embedding spaces and model generalization for dense vision tasks.
• The method shows strong potential for real-world applications with scarce labeled data.