Smoothness- transferred random field

Abstract

We propose a new random field (RF) model, smoothness-transfer random field (ST-RF) model, for image modeling. In the objective function of RF models, smoothness energy is defined with compatibility function to capture the relationship between neighboring local regions, while data energy for the evidence from local regions. Usually, the smoothness energy is constructed in terms of a fixed set of filters or basis which can be learned from training examples and steered to local structures in test examples. ST-RF, on the other hand, takes the data-driven approach to nonparametrically model the compatibility function for smoothness energy. The pipeline for our ST-RF model is as follows: first for each training example, we build a RF model with "ground truth smoothness energy", where the compatibility function is constructed from ground truth value. Then, for each test example, we use data-driven method to find its correspondence with training examples. Lastly, we construct the smoothness energy of ST-RF for each test example by transferring the compatibility function from matched region. After construction, we applies traditional RF inference and learning algorithms to obtain the final estimation. We demonstrate that with transferred ground truth smoothness, random field can achieve state-of-the-art results in stereo matching and image denoising on standard benchmark dataset.