Nguyen,M.H., Wuensche,B., Delmas,P., Lutteroth,C.
3D models have become an essential part of many applications ranging from computer games to architectural design, virtual heritage, and visual impact studies. This has resulted in a critical need for tools which allow non-expert users to easily and efficiently create complex 3D scenes. To answer to that demand, a number of commercial image-based modelling packages have been introduced recently. Such software offers a very intuitive means to create 3D models from a sequence of images. This paper evaluates the most promising 3D reconstruction software packages with regard to efficiency, accuracy, limitations, constraints and compares them with a system developed by us in order to give an insight into their performance. To achieve that, we first describe our own 3D reconstruction system as a reference in order to make deductions about common concepts and differences. Then, we use a set of benchmark datasets to evaluate all considered systems, and gage their limitations with regard to the number of input images they need and the image resolution. Our evaluation shows that as the number of input images decreases, the geometry of models created using correspondence-based approaches contains more holes. However, the structure and geometry still reflect the original model. In contrast, silhouette-based methods produce coarse and distorted geometry as the number of input images decreases. Models obtained using silhouette-based methods from few input images are often unrecognizable.