long-tailed / class-imbalanced classification

Real-world class frequencies are rarely uniform: a few 'head' classes have abundant examples while a long 'tail' of rare classes have very few. Long-tailed recognition is image classification under this heavily skewed distribution, where a naive model trained by plain empirical risk minimisation becomes biased toward the head, achieving high overall accuracy by doing well on common classes while almost ignoring rare ones.

The core difficulty is that the loss is dominated by frequent classes, and the learned classifier weights for rare classes end up with small norms and few updates, so their decision regions shrink. Measuring only overall accuracy hides the failure, which is why long-tailed work reports per-class or class-balanced accuracy and often splits results into many-shot, medium-shot, and few-shot groups.

Remedies fall into several families: re-sampling the data by oversampling tail classes or undersampling head classes; re-weighting the loss, for example by inverse frequency or the 'effective number of samples'; logit adjustment that shifts decision boundaries by the log of the class priors; and two-stage decoupling that first learns features on the natural distribution and then re-balances only the classifier. The unifying idea is to counteract the prior so rare classes get a fair share of the decision space.