ArtClass v2 offers a more realistic, multi-label, and challenging benchmark for artwork FGVC. Our experiments show that current models struggle with hard pairs and multi-label style attribution. Future work includes:
Three art history graduate students per image, with majority-vote labels. Inter-annotator agreement: Fleiss’ κ = 0.81 (style), 0.74 (subject).
: Advanced classes may cover 14 distinct types of paper crafts , including origami (folding) and kirigami (cutting/folding).
[Your Name(s)], [Affiliation(s)]
| Model | Top-1 (std) | Top-1 (hard) | mAP (multi-label) | |------------------|-------------|--------------|-------------------| | ResNet-50 | 62.1% | 41.3% | 0.534 | | ViT-B/16 | 68.3% | 48.7% | 0.612 | | Swin-T | 67.5% | 47.9% | 0.604 | | ConvNeXt-B | | 49.2% | 0.615 |
Art Class has undergone several major iterations, with "v2" marking a significant point in its development when it transitioned toward a more structured web framework.
ArtClass v2 offers a more realistic, multi-label, and challenging benchmark for artwork FGVC. Our experiments show that current models struggle with hard pairs and multi-label style attribution. Future work includes:
Three art history graduate students per image, with majority-vote labels. Inter-annotator agreement: Fleiss’ κ = 0.81 (style), 0.74 (subject). artclass v2
: Advanced classes may cover 14 distinct types of paper crafts , including origami (folding) and kirigami (cutting/folding). ArtClass v2 offers a more realistic, multi-label, and
[Your Name(s)], [Affiliation(s)]
| Model | Top-1 (std) | Top-1 (hard) | mAP (multi-label) | |------------------|-------------|--------------|-------------------| | ResNet-50 | 62.1% | 41.3% | 0.534 | | ViT-B/16 | 68.3% | 48.7% | 0.612 | | Swin-T | 67.5% | 47.9% | 0.604 | | ConvNeXt-B | | 49.2% | 0.615 | Inter-annotator agreement: Fleiss’ κ = 0
Art Class has undergone several major iterations, with "v2" marking a significant point in its development when it transitioned toward a more structured web framework.