Cross-Level Multi-Instance Distillation for Self-Supervised Fine-Grained Visual Categorization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• High-quality annotation of fine-grained visual categories requires extensive expert knowledge, proving time-consuming and costly.
• Existing self-supervised learning methods struggle with fine-grained visual representation due to class-agnostic, patch-level embeddings that overlook critical image regions.

Purpose of the Study:

• To address the limitations of current self-supervised learning for fine-grained visual categories.
• To propose a novel framework, Cross-level Multi-instance Distillation (CMD), that effectively utilizes informative image patches for improved representation learning.

Main Methods:

• The proposed Cross-level Multi-instance Distillation (CMD) framework leverages multiple instance learning to identify and weight important image patches.
• CMD employs both intra-level (within teacher/student nets) and inter-level (between teacher and student nets) multi-instance knowledge distillation on region/image crop pairs.
• This approach comprehensively learns the relationship between informative patches and fine-grained semantic information.

Main Results:

• Experiments on CUB-200-2011, Stanford Cars, and FGVC Aircraft datasets show significant performance improvements.
• The CMD method achieved up to 10.14% higher top-1 accuracy and Rank-1 retrieval compared to contemporary methods.
• CMD surpassed state-of-the-art self-supervised learning approaches by up to 19.78%.

Conclusions:

• Cross-level Multi-instance Distillation (CMD) offers a superior approach for self-supervised learning of fine-grained visual representations.
• The framework effectively addresses the challenge of class-agnostic representations by focusing on discriminative image regions.
• CMD demonstrates strong performance gains, making it a promising solution for fine-grained visual recognition tasks.