Label-free element segmentation of Greek Doric temples from procedurally generated training data

Results record for point-cloud segmentation of classical Doric temples into structuralelements, where the model is trained only on procedurally generated synthetic temples and noannotated real scan enters training. On the published Metaponto benchmark (Pepe et al. 2022, CC BY 4.0), scored over the fiveclasses used by the benchmark authors, the label-free pipeline reaches mIoU 0.8359, against0.8560 for the same model fine-tuned on the benchmark labels and 0.6029 for the authors'random forest on seven geometric features. Input is normals only; absolute elevation is notused. On a second site, the Temple of Apollo at Ancient Corinth, no ground-truth labels exist,so no score is reported; measurements taken from the prediction alone agree with publisheddimensions to within about 2 percent. This deposit is made to record the date on which these results were obtained. It contains theevaluation outputs as recorded, together with summary figures, two short result videos, and aninteractive viewer of the Metaponto result. The pipeline source code and training configurationsare not included. Raw third-party point clouds are not redistributed.