Abstract: Highlights•The proposed method reduces the difficulty of optimizing the intermediate layers.•It bridges the domain and semantic gaps by aligning sketch and image features.•It overcomes the limitation of batch size in metric learning.
Loading