Transfer Learning and Domain Adaptation Techniques in 3D Cultural Heritage Semantic Segmentation: A Systematic Review

Abstract

This systematic review offers an in-depth and critical examination of transfer learning and domain adaptation strategies as applied to the semantic segmentation of 3D cultural heritage data. The inherent complexity, variability, and heterogeneity of 3D heritage datasets often hinder the effectiveness of conventional segmentation models in terms of precision and generalization. To address these limitations, transfer learning and domain adaptation have gained prominence by enabling the reuse of knowledge from pre-trained models and mitigating domain shifts between source and target data. This review systematically categorizes and evaluates existing methodologies, with a particular emphasis on their application within the domain of cultural heritage. It encompasses a broad spectrum of topics, including 3D data acquisition techniques, the design and refinement of deep learning architectures, and the metrics employed for performance evaluation. Furthermore, the review discusses persistent challenges such as limited labeled data, variability in lighting conditions, material textures, and geometric scales. By integrating recent advances and synthesizing key contributions across the literature, this work aims to provide a foundational reference for researchers and practitioners seeking to improve the semantic segmentation of 3D cultural heritage objects and environments.