Fig. 1

The architecture of the Siamese CNN. The network takes a pair of images as input. The original image is sent to one CNN channel while the positive image (from the same scene class) or the negative one (from the different scene class) is sent to the other CNN channel simultaneously. The two parameter-sharing CNNs are exploited to process the input images and extract their features. The nonparametric square layer is used for similarity estimation based on the extracted features of the input pair. At the end of the architecture combines the identification (classification) loss and verification (similarity) loss