Artist-Disjoint Evaluation of Frozen Visual Representations for Art Movement Classification: Pretraining Type, Artist Leakage, and Variability
Year: 2026
Type: article
Abstract: Art movement classification is commonly evaluated on random splits in which works by the same artist appear in both training and test sets, and models are compared on a single split. This study applies an artist-disjoint evaluation protocol to ten frozen visual representations on 65,998 WikiArt paintings by 936 artists in 20 Western movements. All configurations share the same classifier network and the same 20 artist-grouped splits and are compared with Nadeau–Bengio corrected paired tests and Holm adjustment. SigLIP2 So400m/16 achieved the highest 10-fold accuracy, 53.4% (95% CI 46.8–60.0%). The largest differences were associated with pretraining type rather than with architecture or resolution: within ConvNeXt-L, image–text pretraining exceeded supervised ImageNet-22k training by 7.0 points, and CLIP exceeded self-supervised DINOv2 by 6.7 points, both significant after adjustment, also without class weights and with a linear classifier. On an external set of 1223 National Gallery of Art paintings, the model ordering was largely preserved and both differences reappeared (8.6 and 11.2 points). A random split that ignores artists inflated the accuracy of CLIP ViT-L/14 by 11.5 points, and the accuracy of SigLIP2 ranged from 45.9% to 65.8% across artist-disjoint folds. Reliable model comparison in art classification requires artist-disjoint, repeated, and paired evaluation.
Source: Applied Sciences
DOI: 10.3390/app162211491
Author: Bahadır Elmas
Institution: Mimar Sinan Güzel Sanatlar Üniversitesi
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Keywords: CLIP, SigLIP, DINOv2, ConvNeXt, paired comparison test
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7236662748
