Coarse Labels Matter: Revisiting the Role of Coarse-Grained Supervision in Fine-Grained Learning

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• High-quality fine-grained annotations are costly, driving research into using coarse labels for fine-grained learning.
• Current methods often relegate coarse labels to an auxiliary role, favoring complex unsupervised techniques.

Purpose of the Study:

• To propose CSer, a framework that maximizes the effectiveness of coarse label information in Coarse-to-Fine learning.
• To address the challenge of balancing fine-grained feature diversity with robust coarse-grained supervision.

Main Methods:

• Implemented a coarse-grained self-distillation strategy to enhance backbone discriminative power by transferring knowledge from the final classifier to intermediate layers.
• Introduced dense supervision on common component features, decoupled via Non-negative Matrix Factorization, to mitigate simplicity bias in embeddings.
• Utilized intra-class sample relationships to dynamically adjust negative sampling in contrastive learning for tailored fine-grained class relationships.

Main Results:

• Demonstrated the effectiveness of the CSer framework across multiple benchmark datasets.
• Achieved state-of-the-art performance, surpassing existing competing methods in fine-grained learning tasks.

Conclusions:

• The CSer framework effectively leverages coarse label information for superior fine-grained learning.
• The proposed methods successfully reconcile feature diversity preservation with strong coarse supervision, leading to improved model performance.