EKDSC: Long-tailed recognition based on expert knowledge distillation for specific categories

In the field of long-tail visual recognition, the imbalance in data distribution leads to a significant performance gap between head and tail classes. Improving the tail-class performance and alleviating the decline in head class are two critical questions. Although many methods have proposed solutions for the former, most of them fall short in the latter. Introducing additional knowledge is a novel view to address the problem, however, how to attain useful knowledge and further transfer the knowledge to the target model is the core. This paper proposes a novel method called Expert Knowledge Distillation for Specific Categories (EKDSC). Firstly, we propose a kind of well-trained teacher model ensuring each expert concentrates on its specialized field while being less affected by other interference. Furthermore, the teacher model including three categories of experts: head, mid, and tail classes, is utilized to distill their specialized knowledge to the student model. Experimental results demonstrate that EKDSC effectively improves the accuracy of tail classes, and mitigates the common decreases of head classes' performance. Our proposed method achieves a high accuracy, exceeding the current state-of-the-art (SOTA) by 1-5 % on benchmark datasets including the small-scale CIFAR-10 LT and CIFAR-100 LT. Furthermore, it demonstrates outstanding performance on large-scale datasets such as ImageNet-LT, iNaturalist 2018, and Places-LT.