Improving Tail-Class Representation with Centroid Contrastive Learning

In vision domain, large-scale natural datasets typically exhibit long-tailed distribution which has large class imbalance between head and tail classes. This poses difficulty in learning good representations for Recent developments have shown model can be learnt by decoupling the training into representation classifier balancing. However, these works pay insufficient consideration on effect learning. this work, we propose interpolative centroid contrastive (ICCL) to improve ICCL interpolates two images from a class-agnostic sampler class-aware sampler, trains such that of image used retrieve centroids both source We demonstrate effectiveness our approach multiple classification benchmarks.