Rocaille Dataset Technical Validation Artifacts

These are the artifacts produced for the technical validation in the paper "A Multimodal Dataset of 18th-Century Prints for Segmentation and Analysis of Rocaille Ornaments" (Hudcovic et al.): We present a multimodal dataset for the computational study of 18th-century Rococo ornamental prints. Rocaille, a characteristic ornament of the period, exhibits a continuously transforming morphology that has historically resisted systematic formal description. The dataset comprises 1,611 high-resolution digitizations from major German collections, accompanied by expert-curated bilingual metadata and commentary describing iconography and ornamental morphology. A subset of 229 images includes pixel-accurate segmentation masks distinguishing the structural volutes and comb-like extensions of the Rocaille. The images, structured metadata, textual descriptions, and segmentation annotations together support computational analyses ranging from visual multimodal retrieval and semantic segmentation to morphological analysis. As technical validation, we train two standard vision backbones for segmentation, showing that the annotations provide a learnable segmentation signal and that pre-training on both external and on our dataset can support downstream learning. Code, data splits, trained model weights, and evaluation outputs are provided to facilitate reproducible use of the resource. The dataset provides a foundation for cross-disciplinary research at the intersection of art history, digital humanities, and computer vision. The artifacts have been produced by using the dataset described in the paper. The dataset can be found here: https://doi.org/10.5281/zenodo.22978082The code with which the artifacts have been produced can be found here: https://github.com/hudo259/Rocaille-Dataset-Technical-Validation-CodeThe paper can be found here: [TODO] The folder "checkpoints" contains the model checkpoints for both ConvNeXtV2 and Swin Transformer V2 in their Tiny variants. More specifically, the folder contains checkpoints from diffusion-based pre-training on the larger unlabeled subset of the dataset as well as checkpoints from the subsequent fine-tuning on the labeled image-mask pairs of the training split of the dataset. The folder "eval" contains all the results from the evaluation of both models on the test set split of the dataset.Because Zenodo doesn't support uploading large files, checkpoints.zip had to be split into smaller parts. The parts can be merged to one coherent ZIP-file again via: Linux / MacOS / BSD: cat checkpoints.zip.part-* > checkpoints.zip Windows: copy /b checkpoints.zip.001+checkpoints.zip.002+checkpoints.zip.003+checkpoints.zip.004+checkpoints.zip.005 checkpoints.zip Alternatively, selecting all parts and opening them together with 7-Zip will also enable reconstruction of the archive.The code with which these artifacts have been produced can be found here: https://github.com/hudo259/Rocaille-Dataset-Technical-Validation-Code