A 3D Point-Cloud Dataset of Prehistoric Pottery Impressions, with Size-Aware Classification and Denoising Baselines

Pottery excavated from Japan's Jōmon-Yayoi period preserves impressions of seeds and insects that offer rare evidence of early agriculture and food storage, yet identifying the species that formed them is labor-intensive and demands scarce expert knowledge. We present a 3D point-cloud dataset of prehistoric pottery impressions, acquired non-destructively by X-ray CT and released as triangulated surface meshes. The dataset pairs clean laboratory replicas with genuine archaeological specimens across five species—four plant seeds and the maize weevil—together with a Noise-Pair subset of archaeological impressions and their expert-cleaned ground truth. Its defining challenges—classes that share shape but differ in archaeologically meaningful size, structured real-world noise, a built-in laboratory-to-archaeological domain gap, and limited, imbalanced samples—make it a demanding benchmark beyond its heritage context. We provide comprehensive baseline results spanning classical shape descriptors, point-cloud classification networks, and denoising methods. Among these, a size-aware baseline that reintroduces absolute bounding-box dimensions raises cross-domain accuracy on the archaeological data from 34.3% to 71.1% (Macro-F1 35.4% to 61.6%), and morphological denoising raises it further to 73.7% (Macro-F1 66.0%). The dataset and code are released publicly at https://github.com/dbms-ku/jomon-yayoi-3d-impressions to support research in computational cultural heritage and 3D point-cloud learning.