Semi-Supervised Learning for FGVC With Out-of-Category Data

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Current fine-grained visual classification (FGVC) methods predominantly rely on fully-supervised learning, requiring extensive expert labels.
• Semi-supervised learning (SSL) offers a promising alternative by utilizing unlabeled data, but existing SSL paradigms struggle with out-of-category unlabeled data in FGVC.
• The effectiveness of current SSL for FGVC is limited by the assumption of in-category unlabeled data.

Purpose of the Study:

• To develop a novel semi-supervised learning (SSL) method for fine-grained visual classification (FGVC) that effectively incorporates out-of-category unlabeled data.
• To leverage the inherent hierarchical structure of fine-grained categories to improve SSL performance.
• To introduce new strategies for inter-sample consistency regularization and pseudo-relation generation within a hierarchical framework.

Main Methods:

• Proposed a novel SSL design specifically for FGVC that utilizes out-of-category data.
• Assumed and leveraged the natural hierarchical structure of fine-grained categories (e.g., phylogenetic trees).
• Optimized SSL by predicting sample relations within the category hierarchy, introducing strategies for consistency regularization and pseudo-relation generation.

Main Results:

• The proposed method demonstrates significant robustness when dealing with out-of-category unlabeled data.
• The approach can be integrated with existing methods, enhancing their performance.
• The combined approach achieves state-of-the-art results in fine-grained visual classification.

Conclusions:

• The novel SSL approach effectively utilizes hierarchical structures to overcome limitations of out-of-category data in FGVC.
• The method offers a robust and adaptable solution for improving FGVC performance with limited labeled data.
• This work advances SSL techniques for complex visual classification tasks, paving the way for more efficient and accurate models.