A Cross-Scenario Eyeglasses Retrieval Framework Via Landmark Detection

According to the potential of the online eyeglasses market and the lack of search method in eyeglasses websites, this paper study a new type of eyeglasses retrieval. Different from traditional retrieval method, we define the input is the real world photos,which is people who wears eyeglasses, retrieves the pure eyeglasses images in our database.We define it is a cross-scenario eyeglasses retrieval problem.In this paper we mainly discuss how to solve the cross-scenario eyeglasses retrieval prob-lem. By studying object retrieval,cross scenario retrieval and landmark detection technology we design and implement a cross-scenario eyeglasses retrieval framework via landmark detection. We study 33 landmarks to describe a pair of eyeglasses instead of image segmen-tation, and construct the automatic eyeglasses landmark detectors via Supervised Descent Method(SDM), and extend SDM in feature extraction and regression method.To solve the cross-scenario noises, we propose a new eyeglasses shape descriptor, and design a hybrid descriptor which combine color,shape,texture features to represent eyeglasses. And finally we use bag-of-words model to vector quantization.Extensive experiments on real data sets confirm the effectiveness and efficiency of our proposed framework.