ENHANCING AUTOMATION OF HERITAGE PROCESSES: GENERATION OF ARTIFICIAL TRAINING DATASETS FROM PHOTOGRAMMETRIC 3D MODELS

Abstract. Nowadays, many efficient technologies have been developed with the aim of collecting digital images and other metric data, greatly optimising the acquisition procedures and techniques. However, processing this data can be onerous and time-consuming, and increasingly often, there is a need to develop new strategies to enhance the level of automation of these processes. Using artificial intelligence, and particularly Convolutional Neural Networks, it is possible to automate processing tasks such as classification and segmentation. However, a significant challenge is represented by the necessity of obtaining sufficient training data to properly train a deep learning model. These datasets are composed of a significant amount of data and need to be annotated, which sometimes represents an onerous and challenging task. Synthetic data can represent an effective solution to this problem, significantly reducing the time and effort required to manually create annotated datasets and can be particularly useful when studying objects characterised by specific features and high complexity, requiring tailored solutions and ad hoc training. The presented research explores the opportunity of using synthetic datasets – generated from photogrammetric 3D models – for deep-learning-based heritage digitisation applications. The use of synthetic data generated from textured 3D models derived from SfM photogrammetric processes is proposed, with the aim of enhancing automatic procedures in the framework of heritage processes.