We present a method for jointly predicting a depth map and intrinsic imagesfrom single-image input. The two tasks are formulated in a synergistic mannerthrough a joint conditional random field (CRF) that is solved using a novelconvolutional neural network (CNN) architecture, called the joint convolutionalneural field (JCNF) model. Tailored to our joint estimation problem, JCNFdiffers from previous CNNs in its sharing of convolutional activations andlayers between networks for each task, its inference in the gradient domainwhere there exists greater correlation between depth and intrinsic images, andthe incorporation of a gradient scale network that learns the confidence ofestimated gradients in order to effectively balance them in the solution. Thisapproach is shown to surpass state-of-the-art methods both on single-imagedepth estimation and on intrinsic image decomposition.
展开▼
