Patent Dataset

We introduce UPAIR, a novel human-annotated benchmark for evaluating multimodal image retrieval in the context of utility patents. UPAIR addresses current limitations by providing: Three-dimensional relevance annotations between patent image pairs: Visual similarity — resemblance in appearance Semantic similarity — functional or conceptual correspondence Part-whole relationships — whether one image is a component of another Large-scale image-text corpus for training vision-language models with contrastive learning.