Semantic Priors for Intrinsic Image Decomposition (Supplementary)

Our selective search features are formed by concatenating various mask values at a particular pixel, weighted by MCG [1] ‘objectness’ score. We do dimensionality reduction on these features using PCA for efficient computation during reflectance formulation. We use dimensionality reduced features in Stage 2 of the framework unlike Stage 1 as the mid-level priors are re recomputed only in this stage in each iteration. We found no significant change in performance due to this during experimentation. Figure 1 shows a few sample sample masks (overlaid over the image for visualization) and the ‘PCA-image’ (formed by reducing the dimensions to 3) for an example image. Note how in the ‘PCA-image’ the regions belonging to the same object get clustered together illustrating how our selective search features, giś encode mid-level semantics.