Clothing retrieval based on image bundled features

How to evaluate the similarity between two clothing images is the core technical problem of image based clothing retrieval which is extremely useful in aiding online clothing shopping. According to the characteristics of clothing, we address this issue by computing the similarities between the bundled features of different clothing images. Each bundled feature consists of the point features (SIFT) which are further quantified into local visual words in a maximally stable extremal region (MSER). Researches show that bundled feature becomes much more discriminative than single feature, while the intrinsic geometric constraint of a bundled feature is still defective. In this paper, we add a geometric constraint by SIFTs distance matrix to improve the discriminative power. SIFTs distance matrix is constructed by the distances between every two point features (SIFT) in a bundled feature; it has its merits of scale invariance and rotation invariance. Thus, we can match the bundled features of two clothing images and calculate their similarity. Experimental results based on the clothing image database show that our approach works well in the situations with large clothing deformation, background exchange and part hidden, etc.