Improving long-tailed classification by disentangled variance transfer

Image classification is very important in the system of internet of things (IoT), and long-tailed distribution data are common in our daily life. Extremely imbalanced classes in long-tailed classification lead to a huge performance gap between training and testing. A number of methods have been proposed to transfer knowledge from head classes to tail classes, which expects to augment semantic information in tail. However, by projecting feature vectors onto classifier vectors, we find that the projection part and the orthogonal part behave differently in testing phase as the number of instances decreases. In order to properly transfer covariance information in long-tailed classification task, we propose a novel class-based covariance transfer method from the perspective of disentangling. Extensive experimental results on CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT and iNaturalist 2018 illustrate the effectiveness of our method, which will further improve the validity of IoT system.