Beyond Uniform Granularity: Benchmarking Mixed-Granularity Open-Set Recognition

Most existing open-set recognition (OSR) benchmarks implicitly assume uniform category granularity across known class labels. In practice, however, known label sets often span multiple levels of category granularity due to differences in annotator expertise and other practical factors, resulting in a mismatch between such benchmarks and real-world recognition settings. To address this mismatch, we introduce Mixed-Granularity Open-Set Recognition (MG-OSR), where known classes may lie at different levels of an underlying category hierarchy while satisfying an antichain constraint, yet the hierarchy itself is not exposed to the model. We establish an MG-OSR benchmark and evaluate a broad range of conventional and vision-language-model-based OSR methods. We find that this mixed setting poses a threefold challenge to existing methods: (i) unseen subclasses of coarse-grained known classes are prone to being rejected as unknown; (ii) unknowns neighboring fine-grained known classes are prone to being accepted as known; and (iii) shared representation learning exhibits cross-granularity interference, reflected in reduced gradient coherence among sibling categories under the same semantic parent. We further show that widely used metrics, including AUROC and OSCR, are unreliable indicators of granularity-dependent recognition quality in mixed-granularity settings, motivating a granularity-aware evaluation metric that we use to systematically reassess existing methods. Our study identifies mixed-granularity label spaces as an overlooked regime of OSR and provides a unified testbed for OSR beyond uniform granularity.