Classifier Retraining Boosts Long-Tailed Recognition Accuracy.

In machine learning, long-tailed recognition problems arise when some classes have many more training examples than others, leading models to perform poorly on rare "tail" classes. Existing methods often modify loss functions or representations to address this imbalance. This research investigates a simpler, two-stage approach called BS-cRT. The first stage involves training a backbone network and a cosine classifier using Balanced Softmax, a technique designed to mitigate class imbalance. After this initial training, the backbone network is frozen, and only the classifier is retrained. This retraining occurs on balanced episodic batches, maintaining the empirical-prior Balanced Softmax objective. Evaluations across multiple long-tailed datasets, including CIFAR-100-LT, ImageNet-LT, and Places-LT, show that this classifier-only retraining step consistently improves few-shot accuracy. For instance, at an imbalance factor of 100, few-shot gains were substantial, demonstrating that a focused retraining of the classifier can effectively reduce errors in tail classes without complex architectural changes. The study also analyzes limitations of other boundary-probe extensions, reinforcing BS-cRT as a practical and strong baseline.