Test‑Time Discovery (TTD) aims to recognize known classes and detect emerging ones from a continuous unlabeled test stream after deployment. Existing approaches rely on visual clustering or prototype comparison and can only assign anonymous labels to discovered categories, making interpretation difficult in open‑world scenarios. We introduce SEGA (Semantic Grounding with Attributes), which uses large language models to generate attribute‑rich descriptions for known classes and builds an attribute‑anchored visual‑semantic space during training. At test time we apply Relational Semantic Grounding (RSG): unknown samples are anchored to similar and dissimilar known‑class prototypes, the LLM is prompted to produce candidate concepts, and these candidates are re‑ranked in the learned semantic space. The selected concept guides confidence‑based test‑time semantic stabilization. By coupling attribute‑level semantic alignment with relational concept grounding, SEGA converts anonymous test‑time clusters into interpretable semantic concepts. Experiments show that SEGA improves baseline TTD performance while offering semantic grounding for open‑world applications.
Review