Decoupling Representation and Classifier for Long-Tailed Recognition

TLDR

The long‑tail distribution of visual data creates class imbalance challenges for deep learning classifiers. This study aims to decouple representation learning from classification and examine how different balancing strategies influence each component in long‑tailed recognition. The authors train representations independently using instance‑balanced sampling and then train a classifier with various balancing methods, systematically evaluating the impact on performance. They find that representation quality is largely unaffected by imbalance, and that strong long‑tailed performance can be achieved by adjusting only the classifier, achieving state‑of‑the‑art results on ImageNet‑LT, Places‑LT, and iNaturalist. Code is available at the provided URL.

Abstract

The long-tail distribution of the visual world poses great challenges for deep learning based classification models on how to handle the class imbalance problem. Existing solutions usually involve class-balancing strategies, e.g., by loss re-weighting, data re-sampling, or transfer learning from head- to tail-classes, but most of them adhere to the scheme of jointly learning representations and classifiers. In this work, we decouple the learning procedure into representation learning and classification, and systematically explore how different balancing strategies affect them for long-tailed recognition. The findings are surprising: (1) data imbalance might not be an issue in learning high-quality representations; (2) with representations learned with the simplest instance-balanced (natural) sampling, it is also possible to achieve strong long-tailed recognition ability by adjusting only the classifier. We conduct extensive experiments and set new state-of-the-art performance on common long-tailed benchmarks like ImageNet-LT, Places-LT and iNaturalist, showing that it is possible to outperform carefully designed losses, sampling strategies, even complex modules with memory, by using a straightforward approach that decouples representation and classification. Our code is available at this https URL.