Восстановление формы объектов при реконструкции керамических

Abstract

When ceramic vessels are reconstructed from fragments, a significant part of the original surface is often missing altogether and has to be statistically completed rather than merely registered. The aim of the study was to experimentally test a method for completing missing geometry that combines a PointNet++ encoder-decoder architecture with an axis-of-symmetry constraint and a learnable confidence matrix assigning a calibrated point-wise reliability estimate to each generated surface point. The material consisted of 47 ceramic fragments obtained from the Shahristan area of the ancient settlement of Kuva (Kuva District, Fergana Region, Uzbekistan). The encoder aggregates local geometric context through a hierarchical Set Abstraction operation; the decoder restores the missing surface by inverse-distance-weighted feature interpolation (Feature Propagation) regularized by an axis-consistency loss function. The confidence matrix combines the posterior uncertainty of the axis estimate, the predictive confidence of the network and the residual registration error of neighboring points. Applying the axis constraint reduced the mean Chamfer distance from 4.82 to 3.15 mm², and using a learnable confidence matrix instead of fixed weights raised the confidence calibration correlation from 0.41 to 0.74.