Label-Agnostic Category Discovery

Yuwei Bian, Shidong Wang, Chunming Li, Haofeng Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026, pp. 7573-7582

We introduce a label-agnostic paradigm for novel category discovery, designed to operate in open-world settings without relying on assumptions about train-test label space structure. Unlike prior approaches that infer categories from raw data or align to known labels, our method retrieves semantically meaningful concepts from a large-scale, ontology-grounded visual lexicon. This lexicon-guided framework enables discovery that is both scalable and semantically coherent. To support this paradigm, we propose Efficient Probabilistic Sampling (EPS) for prototype-level semantic querying, contrastive representation learning for instance and category discrimination, Adaptive Classifier Assembly (ACA) for dynamic classifier construction, and a hierarchical prototype-centroid alignment strategy for estimating category count. Taken together, these components instantiate Label-Agnostic Category Discovery (LACD) as a practical and principled solution for open-world discovery with explicit granularity control. Extensive experiments on standard benchmarks demonstrate that LACD exhibits strong clustering performance on specific unlabeled datasets when supported by a lexicon.