Read the original document by opening this link in a new tab.
Table of Contents
1. Introduction
2. Related Work
3. Parametric Contrastive Learning
4. Experiments
Summary
Parametric Contrastive Learning (PaCo) is proposed to tackle long-tailed recognition by introducing parametric class-wise learnable centers to rebalance from an optimization perspective. The paper analyzes the PaCo loss under a balanced setting and shows adaptive enhancement in pushing samples close to their corresponding centers. Experiments on various benchmarks demonstrate state-of-the-art results for long-tailed recognition. The study also explores rebalancing in supervised contrastive learning, highlighting the importance of low-frequency classes. Key contributions include identifying the shortcomings of supervised contrastive learning in imbalanced settings, extending it to PaCo loss with learnable centers, and achieving new records for long-tailed recognition.