A Probabilistic Segmentation Scheme

Joint Image GMM and Shading MAP Estimation

“…2D may require lots of iterations to converge. Alternatively, to recover a better shading, a segmentation should be introduced into the model, similar to [1]. Different objects in the image could then have different GMM models and shadings, and the sharp boundary would be ensured by the segmentation.…”

“…K) and initial shading h = 0. The algorithm iteratively updates variables (α, π k , Σ k ), via (9), (10), (12), (13), and variables (h, µ) via solving linear system (11), (1). Each of the updates maximizes the objective E(h, θ, α) w.r.t.…”

“…This work proposes a novel model to estimate shading in natural images 1 . We assume some areas of the image are brighter or darker because of the illumination/shadows.…”

“…This work was mostly inspired by [1], where the segmentation model includes segmentation and shading. In the case of MAP shading recovery, the estimation of the shading (by a discrete optimization) is alternated with estimation of all appearance parameters, which is inferior to our method.…”

Combining Shape Priors and MRF-Segmentation

“…The number of segments and the numbers of Gaussians for each segment (typically from 2 to 4) as well as the parameters α and λ were estimated experimentally. The appearance model includes additional a-priori slowly varying shading fields for each segment (see [13] for details). These shadings were learned simultaneously withq in a fully unsupervised manner.…”