Artclass — V2

ArtClass v2: A Benchmark for Fine-Grained Artwork Attribute Recognition with Multi-Label Stylistic Labels Authors [Your Name(s)], [Affiliation(s)] Abstract Fine-grained visual classification (FGVC) of artwork is challenging due to high intra-class variance, subtle inter-class differences, and domain-specific attributes (e.g., brushwork, palette, era). We introduce ArtClass v2 , a new benchmark dataset consisting of 120,000 labeled artwork images spanning 150 artist styles, 12 historical periods, and 8 medium types (oil, watercolor, etc.). Unlike its predecessor, ArtClass v2 provides multi-label annotations (style + period + subject matter) and is designed to handle real-world art collection scenarios with class imbalance and partial labels. We evaluate 10 state-of-the-art FGVC architectures (e.g., DenseNet, Vision Transformers, MLP-Mixers) and show that even top models achieve only 68.3% top-1 accuracy, leaving significant room for improvement. ArtClass v2 is publicly available to spur research in computational art history and digital humanities. 1. Introduction The intersection of computer vision and art history has grown rapidly, enabling tasks such as artist attribution, style classification, and digital cataloging. Early benchmarks like the ArtClass v1 dataset provided a foundational 50-class artist classification task [1]. However, real-world art collections present more nuanced challenges: an artwork may belong to multiple overlapping styles (e.g., “Impressionism” and “Landscape”), span multiple temporal categories, or include ambiguous attributions.