Structure-from-Motion solutions for heritage 3D documentation: pros and cons, issues and solutions

Abstract

Since their introduction in the heritage field about fifteen years ago, automated image processing techniques known as Structure from Motion (SfM) and dense Multi-View Stereo (MVS) have become very popular in archaeology and the wider field of cultural heritage. Nowadays, SfM and MVS pipelines can almost be considered standard tools in many aspects of cultural heritage documentation. Although many papers have showcased how SfM-based processing chains can be used to transform 2D photographs into map-like orthophotographs and textured 3D models, the geometric and radiometric accuracy of the output is seldom thoroughly assessed. Additionally, the wide variety of available software packages and their tunable parameters often produce very dissimilar results from the same input, notwithstanding reliability and repeatability issues. This tutorial will therefore provide a critical assessment of SfM-based approaches aimed at cultural heritage documentation. Since it aims at young researchers and less-experienced practitioners in this field, the tutorial will start with a theoretical underpinning of image-based modelling and the algorithmic steps included in most SfM solutions. Afterwards, using several terrestrial and airborne image sets covering various archaeological sites, the SfM-based modelling pipelines will be illustrated with their possibilities and pitfalls, answering questions like: “What are the accuracies one can expect?”, “Which input variables are essential for the method ‘to properly work?”, “In which case is a SfM approach (un)suited to document sites or artefacts?”, “Do specific ‘tricks’ exist to make SfM work?”. In the end, it is hoped that the participants will have a more thorough understanding of SfM-based workflows so that they will be able to successfully apply such approaches in the field of cultural heritage on the one hand, and critically validate the generated output on the other. Since the imagery will be processed with several commercial and freeware packages, the participants should bring their own laptop (preferably a 64-bit Windows machine with at least 6/8 GB of RAM and, if possible, an OpenCL-capable graphics card).