Part-Guided Relational Transformers for Fine-Grained Visual Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Convolutional Neural Networks (CNNs) have advanced fine-grained visual recognition.
• Distinguishing between visually similar subcategories remains a significant challenge.

Purpose of the Study:

• To propose a unified framework, PArt-guided Relational Transformers (PART), for fine-grained visual recognition.
• To address the challenge of subtle differences by constructing feature-level interrelationships and capturing part-level discriminative features.

Main Methods:

• Developed PART, a framework adapting Transformer models for visual recognition.
• Implemented an automatic part discovery module to identify discriminative regions.
• Utilized a feature transformation module to explore correlations between global and part embeddings.

Main Results:

• Achieved state-of-the-art performance on three widely-used fine-grained object recognition benchmarks.
• Demonstrated the effectiveness of the approach through experimental results and visualizations.
• The framework does not require additional part branches during inference.

Conclusions:

• The PART framework effectively enhances fine-grained visual recognition by integrating part-level information and feature relationships.
• The proposed method offers an efficient and effective solution for classifying objects with subtle visual variations.
• Explainable visualizations confirm the model's ability to focus on discriminative object parts.