Class Distribution Pyramid Experts for Long-tailed Visual Recognition

Deep long-tailed visual recognition is a challenging task due to its extreme imbalance in classes. Existing methods improve the overall performance by making a trade-off between head and tail classes. We conduct experiment find that in this way, not only the performance of the head classes is compromised, but also the performance of the tail classes is not maximized. Based on these observations, we propose Class Distribution Pyramid Experts(CDPE), a clean and simple framework for end-to-end training.T each expert are calculated by insta hrough the idea of divide and conquer, we maximize the representation of classes at different quantitative levels by different experts and further propose a novel instance-aware aggregation network that dynamically integrates the predictions of different experts. We conducted extensive experiments on CIFAR10/100-LT, ImageNet-LT and iNaturalist2018 datasets to demonstrate our method’s effectiveness, practicability, and state-of-the-art performance. We hope this simple method will motivate people to rethink the divide-and-conquer strategy in long-tailed visual recognition.