A Unified Objective for Novel Class Discovery

Generalized Category Discovery

“…For the baselines from NCD, we follow the original implementations and learning schedules as far as possible, referring to the original papers for details [12,16]. For these experiments, we freeze the backbone and train only the classification heads.…”

“…To the best of our knowledge, there is no baseline for GCD that can be used as-is from the literature. Instead, we adapt two NCD baselines for it: RankStats [16], which is widely used as a competitive baseline for novel category discovery, and UNO [12], which is to the best of our knowledge the state-of-the-art method for NCD.…”

“…Jia et al [21] propose an end-to-end NCD method for singleand multi-modal data with contrastive learning and winnertakes-all hashing. A unified cross-entropy loss is introduced in UNO [12] to allow the model to be trained on labelled and unlabelled data jointly by swapping the pseudo-labels from labelled and unlabelled classification heads.…”

“…Next, we develop a method to estimate the number of categories in the unlabelled data, which is a challenging task that is understudied in the literature. Finally, we build two strong baselines for this problem by modifying stateof-the-art NCD methods, RankStats [16] and UNO [12], to fit with our setting.…”

Novel Class Discovery without Forgetting

“…NCL [61] introduces contrastive learning and OpenMix [62] uses a convex combination of labeled and unlabeled instances to enhance class discovery. UNO [16] learns a unified classifier which identifies labeled and unlabeled instances using groundtruth labels and pseudo-labels respectively. Joseph et al [28] uses cues from multi-dimensional scaling to enforce latent space separability, while Jia et al [26] leverages contrastive learning with WTA hashing to enhance class discovery.…”