classifier-retraining

This paper proposes BS-cRT, a two-stage baseline for long-tailed recognition that trains a backbone with Balanced Softmax and then retrains only the classifier on balanced batches, achieving consistent few-shot accuracy improvements on multiple benchmarks.