Deep Spatial-Semantic Attention for Fine-Grained Sketch-Based Image Retrieval

Here we offer a more detailed description of the proposed model to facilitate re-implementation. A schematic illustration of the network architecture of the proposed model can be found in Fig. 2 of the main paper. It shows that the model is a Siamese triplet network with three branches of identical architectures and shared parameters. In this section, we further describe the detailed network architecture for each branch. Table 1 shows that each network branch has 7 convolutional layers and 2 fully connected layers. The first 5 convolutional layers are the same as the Sketch-a-Net [5] and the last 2 convolutional layers are part of the proposed attention module. Note that the output of the attention module is a 7× 7 attention mask which is used to re-weight the feature map of pool5. The attention module shortcut connection (see Sec. 3.2 of the main paper) takes place before Layer No. 12 and the coarse-fine fusion shortcut connection occurs before Layer No. 14 (as detailed in Sec. 3.3 of the main paper). The final output of each branch is a 512D feature vector which is then subjected to our HOLEF loss.