Recognizing Indoor Images with Unsupervised Segmentation and Graph Matching

In this paper we address the problem of recognizing scenes by performing unsupervised segmentation followed by matching the resulting adjacency region graph. Our segmentation method is an adaptive extension of the Asymetric Clustering Model, a distributional clustering method based on the EM algorithm, whereas our matching proposal consists of embodying the Graduated Assignement cost function in a Comb Algorithm modified to perform constrained optimization in a discrete space. We present both segmentation and matching results that support our initial claim indicating that such an strategy provides both class discrimination and individual-within-a-class discrimination in indoor images which usually exhibit a high degree of perceptual ambiguity.