contempArt embeddings: 16 image-embedding heads across 9 backbones
收藏Zenodo2026-04-21 更新2026-05-26 收录
下载链接：
https://zenodo.org/doi/10.5281/zenodo.19685513
下载链接
链接失效反馈官方服务：
资源简介：
Precomputed image embeddings for the contempArt dataset (441 contemporary German art students, 14,390 artworks). Ships 16 embedding heads across 9 backbones: CLIP (ViT-L/14, ViT-B/32), OpenCLIP (ViT-L/14, ViT-H/14 via KIM-C), DINOv2 (style via fruit-SALAD, gram, CLS, layer-12), CSD (style + content), SCFlow (style + content), GOYA (style + content), SD 2.1 VAE (KIM-A appearance), and SSCD (near-duplicate).
Each .npz is keyed by a filenames array plus one or more embedding arrays. All rows are aligned across files. See the README.md inside the record for full format, head table, and quickstart.
Derived from the original contempArt dataset (Huckle, Garcia & Nakashima, ECCV 2020 workshop). Analysis code and results: github.com/georgeblck/contempart-reanalysis.
License: CC-BY-NC 4.0, chosen conservatively because SSCD weights carry a non-commercial clause that propagates to derived features. Individual backbones are more permissive; see the record README for per-model licenses.
提供机构：
Zenodo创建时间：
2026-04-21
