Single Image Intrinsic Decomposition Using Transfer Learning

Abstract

Separating an RGB image into its albedo and shading maps abstracts the idea behind intrinsic image decomposition. It is a prolonged and underdetermined problem in computer vision, and we aim to tackle this problem using an efficient and optimized method. This work is an attempt to decompose the input image into its albedo and shading component using a Deep Convolution Neural Network (DCNN) model considering an encoder-decoder based architecture. Our model uses a pre-trained DenseNet-169 as an encoder for the input image and two decoders for the two output images viz. albedo and shading. The synthetic ground truth images from MPI-Sintel dataset helps in the dense supervision of our model. We show results for both synthetic images and real images from the IIW dataset. Our architecture aims to reduce the number of parameters used while training and has produced better results when compared with previous approaches on the MPI-Sintel dataset.