Class-aware Universum Inspired re-balance learning for long-tailed recognition

Open Webpage

Enhao Zhang, Chuanxing Geng, Songcan Chen

Published: 01 Jan 2025, Last Modified: 26 Sept 2025Pattern Recognit. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•We address the LTR from the Universum’s perspective by innovatively endowing Universum with class-aware ability, thereby proposing a novel rebalancing learning strategy.•To select appropriate Universum, we introduce the C2Gand devise a HoMu technique to autonomously generate CaU without relying on external data.•We verify both theoretically and experimentally the excellent performance of CaUIRL, especially generalization on minority classes.