LLM meets Vision-Language Models for Zero-Shot One-Class Classification

• DOI: 10.48550/arxiv.2404.00675
• Publication Date: 2024-03-31
• Authors: Yassir Bendou; Giulia Lioi; Bastien Pasdeloup; Lukas Mauch; Ghouthi Boukli Ha...; Fabien Cardinaux; Vincent Gripon
• Keywords: Zero (linguistics)

We consider the problem of zero-shot one-class visual classification. In this setting, only label target class is available, and goal to discriminate between positive negative query samples without requiring any validation example from task. propose a two-step solution that first queries large language models for visually confusing objects then relies on vision-language pre-trained (e.g., CLIP) perform By adapting large-scale vision benchmarks, we demonstrate ability proposed method outperform adapted off-the-shelf alternatives in setting. Namely, realistic benchmark where are drawn same original dataset as ones, including granularity-controlled version iNaturalist, at fixed distance taxonomy tree ones. Our work shows it possible single category other semantically related ones using its