Illumination-invariant Statistical Shape Recovery with Contiguous Occlusion

Spherical harmonics (SH) has been an attractive fit for illumination modeling in shape recovery after the conclusion drawn by Basri and Jacobs. The main challenge is the computation of the spherical harmonics projection (SHP) images to be robust against imaging conditions other than illumination. Occlusions due to wearing apparel and makeup, or even incompliance to the requirement of the statistical model introduce errors in the reconstructed SHP images which in turn has a direct impact on the recovered shape. In this paper, we propose to cast errors introduced due to occlusion as: (1) statistical outliers which are determined and rejected using robust statistics and (2) local spatial erroneous continuous regions where Markov Gibbs random field with the homogenous isotropic Potts model is adopted to model the occlusion's spatial interaction. Our results show the effectiveness of the proposed algorithms in handling high levels of contiguous occlusion compared to one of the state-of-the-art statistical illumination invariant shape-from-shading. In particular, MGRF and robust estimation using Geman-McClure function outperform the singular value decomposition (SVD) performance approach which is very sensitive to the presence of occlusion even at low levels. In the meantime, the performance of Lorenztian function approaches SVD due to the presence of errors caused by basis of different identity than the shape to be reconstructed. We provide empirical validation of our conclusions by simulations and real experiments.