Kernel likelihood estimation for superpixel image parsing

Abstract

In superpixel-based image parsing, the image is first segmented into visually consistent small regions, i.e. superpixels; then superpixels are parsed into different categories. SuperParsing algorithm provides an elegant nonparametric solution to this problem without any need for classifier training. Superpixels are labeled based on the likelihood ratios that are computed from class conditional density estimates of feature vectors. In this paper, local kernel density estimation is proposed to improve the estimation of likelihood ratios and hence the labeling accuracy. By optimizing kernel bandwidths for each feature vector, feature densities are better estimated especially when the set of training samples is sparse. The proposed method is tested on the SIFT Flow dataset consisting of 2,688 images and 33 labels, and is shown to outperform SuperParsing and some of its extended versions in terms of classification accuracy.