Computer Vision News - September 2019

Intrinsic Image Decomposition (IID) 5 Method The method composed of three building blocks: semantic feature selection, shading formulation, and reflectance formulation. Semantic Features: The semantic features component produces two sets of feature vectors. The first set is feature vectors generated for each 30x30 (non-overlapping) patches. Each patch is passed through Region-based Convolutional Neural network (RCNN) where the last fully connected layer is defined to be the feature vector f b for each of the patches in the image. The second set of features are the selective search features. These features give interesting image regions that have a high probability to contain an object. For each detection proposal, the method generates dense binary region masks and score. Each pixel is assigned with a feature vector g i of proposal masks weighted by proposal score. Shading Formulation: The authors suggest using the RCNN and selective search features to define pixel neighborhoods. At each iteration of the method, the shading and reflectance are estimated by minimizing the following energy function: Here S g , S m , S l are the global, mid-level and local shading priors, each weighted differently by a different coefficient. The global shading is further divided into sparse neighborhood consistency term S c and weight propa- gation term S p such that: S g =S c +S P . We next explain the terms in the energy function: 1) S c is the sparse neighborhood consistency. It is computed by summing the squared differences between each pixel and a weighted average of its neighbors' pixel (where the neighbors are defined by b ).