3D object reconstruction using multiple views

Abstract

3D object modelling from multiple view images has recently been of increasing interest in computer vision. Two techniques, Visual Hull and Photo Hull, have been extensively studied in the hope of developing 3D shape from multiple views. These early methods have the advantage that they do not require pre-processing procedures such as feature selection and matching, which fail when images are of low resolution. One drawback of these two methods is their discrete formulation, which is demanding of memory and limits the type of optimisation methods that can be used. This study proposes a continuous formulation in contrast to the discrete formulations typical of these earlier methods, and aims to robustly reconstruct the 3D shape and colour of an object seen in a multi-view system. The use of a continuous formulation based on kernel density estimates enables us to define a gradient ascent algorithm (e.g. a mean shift algorithm) to recover the 3D shape and colour. Moreover, we propose to include prior information in this continuous modelling to improve the quality of the reconstruction. The proposed approach has several advantages: it is less memory demanding, the resulting algorithm is suitable for parallel processing, and it recovers concavities that are usually lost when estimating shape from silhouettes with the standard visual hull method.