Exploiting Supervised Learning for 3D Model Semantic Segmentation Using Multispectral Data

Abstract

3D model texture-based segmentation using multispectral imagery to define its construction materials is addressed within the scope of this work. An end-to-end pipeline is proposed to digitize a real-world object, construct a spatial consistent multispectral texture map and to identify materials on its surface. A multispectral camera capable of capturing ultraviolet to near infrared imagery is used to create image sequences for its Structure-from-Motion based 3D reconstruction. We utilize computational geometry techniques to create a spatial-consistent texture based on ultraviolet to near infrared imagery. Various supervised learning approaches are utilized and evaluated on the identification of materials on a 3D model's surface. Experimental results are promising and reveal its capabilities in the study of 3D digitized models.