Abstract
While many existing modelling techniques achieve impressive results, they are often reliant on assumptions such as prior knowledge of the scene's surface reflectance. This considerably restricts the range of scenes that can be reconstructed, as these assumptions are often violated in practice. One technique that allows surface reconstruction regardless of the scene's reflectance model is Helmholtz Stereopsis (HS). However, to date, research on HS has mostly been limited to 2.5D scene reconstruction. In this paper, a framework is introduced to perform full 3D HS using Markov Random Field (MRF) optimisation for the first time. The paper introduces two complementary techniques. The first approach computes multiple 2.5D reconstructions from a small number of viewpoints and fuses these together to obtain a complete model, while the second approach directly reasons in the 3D domain by performing a volumetric MRF optimisation. Both approaches are based on optimising an energy function combining an HS confidence measure and normal consistency across the reconstructed surface. The two approaches are evaluated on both synthetic and real scenes, measuring the accuracy and completeness obtained. Further, the effect of noise on modelling accuracy is experimentally evaluated on the synthetic dataset. Both techniques achieve sub-millimetre accuracy and exhibit robustness to noise. In particular, the method based on full 3D optimisation is shown to significantly outperform the other approach.