FGSR: A Fine-Grained Ship Retrieval Dataset and Method in Smart Cities

Figure 3

The framework of our fine-grained ship reidentification network. Firstly, we input the ship image into a convolutional network to obtain a pyramid feature map. In pyramid feature maps, all levels are semantically strong, including the high-resolution levels. Then, we apply a pyramid fusion module to aggregate all scare feature maps. To identify the moveable facilities on the ship, we design an occlusion module to estimate an occlusion map which identifies the areas that can be changed in other time slots. After that, we combine the occlusion map and the aggregated feature to produce an occlusive feature. Finally, the multibranch identity module takes the occlusive feature as input to produce the identity feature vector for the corresponding image.