Attribute-Assisted Domain Transfer from Image to Sketch

Data-inefficiency greatly hinders sketch-related applications and studies. Transferring knowledge from the realimage domain to the sketch domain may eliminate this obstacle. However, a huge domain gap exists between images and sketches. To reduce the domain shift between the image and the sketch, we propose an attribute-assisted domain transfer method. By separating the shared geometric features from the private semantic features of the real images, the proposed attribute-assisted networks (ASN) can learn more effective domain-invariant features for unsupervised domain adaptation. Extensive results on the Imageto- Sketch task demonstrate the effectiveness of the proposed method.