Emery Pierson, , Diego Gomez, Souhail Hadgi, Maks Ovsjanikov, Diego Jimenez-Badillo

Archaeological collections increasingly include 3D digitizations, but finding stylistically related objects still depends on expert knowledge and manual cataloguing. We evaluate pretrained vision and 3D foundation models as annotation-free descriptors for 84 pre-Hispanic stone masks with expert-assigned styles from the Museo del Templo Mayor in Mexico City. We compare image and point-cloud embeddings from photographs, grayscale and normal-shaded renderings, and sampled 3D geometry. We evaluate the embeddings on downstream tasks: retrieval, classification, and unsupervised clustering, where foundation embeddings outperform traditional handcrafted descriptors. Our results suggest that geometric cues and semantic supervision of features help to distinguish those artifact styles.

More details to come!!

Citation

If you consider our work useful, please cite:

@misc{incoming}

This webpage was inspired by Nerfies.