Contour detection network for zero-shot sketch-based image retrieval

Abstract Zero-shot sketch-based image retrieval (ZS-SBIR) is a challenging task that involves searching natural images related to given hand-drawn sketch under the zero-shot scene. The previous approach projected and features into low-dimensional common space for retrieval, used semantic transfer knowledge of seen unseen classes. However, it not effective enough align multimodal when projecting them space, since styles contents sketches are different they one-to-one correspondence. To solve this problem, we propose novel three-branch joint training network with contour detection (called CDNNet) ZS-SBIR task, which uses maps as bridge alleviate domain gap. Specifically, use metrics constrain relationship between sketches, so can be aligned in space. Meanwhile, further employ second-order attention capture target subject information increase performance descriptors. In addition, teacher model word embedding method Extensive experiments on two large-scale datasets demonstrate our proposed outperforms state-of-the-art CNN-based models: improves by 2.6% Sketchy 1.2% TU-Berlin terms mAP.