Object localization with discriminate superpixels

We propose a simple image component parsing method based on discriminate regions. We define a discriminate region as some image areas which can be easily identified as foreground or background. For instance, eyes are the discriminate regions for face because it can hardly be found in background, while trees are discriminate ones as well since we can hardly found similar regions in face. Foreground should contain more discriminate regions than that of background. Based on this idea, we employ superpixels to mine those discriminate regions and construct a graph to reveal the relationship among them. We compute the unary and pairwise potential of the graph and then carrying out inference with conditional random field model (CRF). Finally pixels could obtain the most possible classes they belong. Our method performs excellently in two image databases: CaltechlOl andGraz02.