CSIR4G: An effective and efficient cross-scenario image retrieval model for glasses

In this work, we propose an effective and efficient cross-scenario glasses retrieval model with the goal of providing a new online shopping experience. Two challenging issues arise for cross-scenario glasses retrieval: identifying glasses in a query image and extracting features from the identified glasses. To properly address these issues, we introduce a novel segmentation-free framework which includes a glasses detection model, an attribute recognition model, and a coarse-to-fine search strategy. Specifically, a new keypoint-based scheme called EyeGlasses keYPoinT (EGYPT) for glasses detection is proposed, in which a set of representative keypoints along a glasses frame are automatically identified. Based on these detected keypoints, various local feature descriptors are extracted to learn the semantic attributes of the glasses. Furthermore, fast retrieval is achieved by applying a hierarchical search strategy based on the index of the attributes. An experimental analysis with real-world photos and product images verifies the efficacy and efficiency of our proposed retrieval model.