UniFGVC

We present a dataset of structured textual descriptions for visual categories, aimed at supporting fine-grained visual recognition and multimodal understanding. The dataset emphasizes category-level, discriminative attributes and semantic relations, rather than instance-specific or generic captions. By organizing descriptions in a structured format, it enables effective use as textual priors for retrieval, reasoning, and zero-shot or few-shot recognition, while reducing redundancy and hallucination. The dataset is model-agnostic and can be seamlessly integrated into various vision\u2013language frameworks, facilitating scalable category extension without additional training. This resource provides a practical benchmark for advancing fine-grained classification and multimodal learning.