Category-agnostic object re-identification

Object re-identification, a fundamental task of pattern recognition, has long been constrained by its reliance on category alignment between query and gallery. To break this limitation, we pioneer Category-Agnostic object Re-Identification (CA-ReID), a brand new task enabling identity retrieval of arbitrary objects (including novel categories unseen during training) from gallery containing mixed categories. To address the dual-granularity challenge of joint inter-category discrimination and intra-category identification involved in this task, we draw inspiration from human cognitive system, where coarse categorization and fine identification share neuronal substrates. Specifically, we propose a unified framework to decouple category-related and identity-related features. For each single-granularity feature learning, we apply the classical cross-entropy and triplet loss, and a customized VLM-based visual-semantic alignment loss, as well as the negative learning loss, for identity-aware and category-aware feature learning. Besides, we propose a novel cross-granularity correction loss for information interaction between dual-granularity features. For this new problem, we build a large-scale benchmark dataset CA-ReIDB, experiments on which demonstrate that the proposed framework outperforms existing state-of-the-art methods, proving its viability as a foundational step toward practical real-world ReID systems. © 2026 Elsevier Ltd