Classifying Decorated Pottery by Reshaping Embedding Geometry

The classification of decorated ceramics, and specifically the majolica from Montelupo Fiorentino, is challenging because visual categories are strongly imbalanced, internally heterogeneous, and separated by subtle decorative differences. This study investigates whether the latent geometry learned by a global convolutional classifier can be reorganized into more coherent local classification problems. ResNet embeddings are represented through class-specific prototypes, which are subsequently grouped into visual regions. Region-specific Siamese networks refine the local embedding geometry, while a learned router and a final gate determine when specialist predictions should replace the global decision. In the fully deployable setting, the proposed pipeline improves the global ResNet baseline from 68.5% to 72.2% accuracy and from 49.3% to 51.5% macro-F1. These results show that local specialization can recover information not fully exploited by the global decision boundary, although its practical benefit remains constrained by region prediction, specialist reliability, and the limited support available for the rarest classes.