Linking Known and Unknown: Generalized Cross-Instance Feature Helps Category Discovery

Abstract

In this paper, we tackle Generalized Category Discovery (GCD) by drawing inspiration from the platypus—a creature that uniquely blends features from different species. Our method bridges the gap between known and unknown categories through a novel cross-instance feature learning paradigm. Unlike traditional GCD methods, we dynamically mix patches from multiple images, integrating features across instances. This process is enhanced by a progressive mixing strategy that evolves from focusing on labeled data to incorporating both labeled and unlabeled data. Complemented by a hierarchical contrastive learning framework that enforces constraints at global, mixed-origin, and inter-mixed levels, our approach effectively generalizes across different feature spaces. Extensive experiments on six benchmark datasets demonstrate our method’s superior performance in recognizing known classes and discovering new ones, setting a new standard in GCD tasks.