Labeling image patches by boosting based median classifier

Abstract

This paper presents a median based classifier which predicts the conditional median of the class label given the feature vector. The class label is defined through a hidden variable whose median is further modeled as an additive model of the feature vector. We propose to estimate the conditional median of the hidden variable given the feature vector in the framework of generic functional gradient algorithm [5]. An equivalent form of the definition of median is introduced, whose smoothed version is employed as the objective function. To fit the model for the conditional median, the proposed objective function is maximized by gradient ascent in functional space, updating the fitted model a small step in the gradient direction in each iteration. The resulting algorithm, Median Boost, is a boosting like procedure which obtains the informative features and the classifier at the same time. On the task of labeling building blocks in natural images, the comparison results show that Median Boost performs better than or similar to several alternatives.

Department(s)

Mathematics

Document Type

Conference Proceeding

DOI

https://doi.org/10.5244/C25.130

Publication Date

1-1-2011

Journal Title

BMVC 2011 - Proceedings of the British Machine Vision Conference 2011

Link to Full Text

Share

COinS