Abstract
-
Support Vector Regression (SVR) has been a long standing
problem in machine learning, and gains its popularity on various
computer vision tasks. In this paper, we propose a structured support
vector regression framework by extending the max-margin principle to
incorporate spatial correlations among neighboring pixels. The objective
function in our framework considers both label information and pairwise
features, helping to achieve better cross-smoothing over neighboring
nodes. With the bundle method, we effectively reduce the number
of constraints and alleviate the adverse effect of outliers, leading to an
efficient and robust learning algorithm. Moreover, we conduct a thorough
analysis for the loss function used in structured regression, and
provide a principled approach for defining proper loss functions and deriving
the corresponding solvers to find the most violated constraint. We
demonstrate that our method outperforms the state-of-the-art regression
approaches on various testbeds of synthetic images and real-world scenes.