Computer Vision News - September 2019

2) S p is the weight propagation term. This term is computed by a weighted sum of the square of differences between each pixel shading value to its neighbor patches. 3) S m is the mid-level context, generated in the same way as S p but here the weights are defined by the selective search features g i to penalize dissimilar pixel that belongs to the same object. 4) S l is the local context, generated by a weighted sum of the log image intensity and the log shading values. This aims to enforce resemblances between the original image and the shading image. Reflectance Formulation: The reflectance formulation enforces color sparsity using L 1 terms. To this end, the authors use a sum of four weighted terms to define the reflectance energy: Here again R g , R m , R l , R a stand for global mid-level local and image approximation term. Each of the first three terms is a weighted sum of the L 1 differences between the reflectance value of a pixel and its neighborhood (the difference between the three is the neighborhood and the weights). The last term R a enforces continuity between the shading and reflectance stages by enforcing the reflectance estimate from the current stage to be similar to the intermediate reflectance solution for the previous shading iteration. These terms connect between the two stage of the optimization and make the method more robust and with fewer artifacts. Below is a figure that summarizes the method: Research 6

RkJQdWJsaXNoZXIy NTc3NzU=