Long-Tailed Recognition Drops the Uniform Collapse Target: Tail Classes Keep Their Spread and Four Benchmarks Beat Prior Methods in Overall Accuracy

Long-tailed recognition no longer asks every class to converge on the same set of fixed feature directions: head classes still tighten up, while tail classes keep their within-class spread. Earlier long-tailed methods took a fixed equiangular tight frame, one shared set of target directions whose pairwise angles are equal, as a uniform alignment target for all classes. Forcing the tail into that uniform target over-contracts its samples and weakens generalization; releasing it gives better overall accuracy than prior methods on CIFAR100-LT, ImageNet-LT, Places-LT and iNaturalist-2018.

The change sits in two coordinated controls on the backbone, one setting how much dimensionality the collapse occupies and the other setting the clearance between classes. A class-adaptive sparse activation classifier uses feature-conditioned gating to specialize nodes per class, so tail classes occupy their own feature directions, and a frame-constrained adaptive margin loss sets the margin by class frequency, small for head classes and larger for tail classes. In the empirical analysis, within-class spread rises monotonically from head to tail, consistent with the tail retaining variability instead of fully collapsing.

The demonstrated scope is the overall-accuracy comparison across four long-tailed image recognition benchmarks, plus the empirical trend of within-class spread rising monotonically from head to tail classes.

The monotonic rise in within-class spread from head to tail is reported as consistent with the tail keeping its variability, an agreement rather than a separate test; that full collapse over-contracts tail samples and weakens generalization is the premise the change rests on.