Pseudo-supervised contrastive learning with inter-class separability for generalized category discovery

Abstract

Generalized Category Discovery (GCD) addresses the task of transferring knowledge from labeled known categories to recognize both known and novel categories in an unlabeled dataset. The key challenge is the fact that novel categories has no prior information. To tackle this problem, we propose a novel method to learn discriminative features for both known and novel categories called pseudo-supervised contrastive learning with inter-class separability (PIS). To expand the distinction between known and novel categories, we employ OpenMax to separate the known and novel data in the unlabeled set. Based on Earth Mover’s Distance, we introduce an inter-class separation loss to explicitly enlarge the distance between known and novel instances. Additionally, in order to expand inter-class differences of the novel categories, we leverage the large pretrained model CLIP to assign pairwise pseudo-labels for novel instances. Then pseudo-supervised contrastive learning is adopted to narrow the distance between pairs of instances with the same pairwise pseudo-labels and enlarge the distance between pairs of instances with different pairwise pseudo-labels. Through comprehensive evaluations on generic image recognition datasets and challenging fine-grained datasets, we show that our PIS achieves state-of-the-art performance.