Fig. 2

The visualization of the joint decision-making mechanism. Each shape represents a scene class. One test image along with the images in the comparison set is fed through the whole network. The identification model outputs the class probability Pi(j) (j = 1, 2, ..., K) of images in the comparison set. Meanwhile, the verification model outputs the similarity measurement Pv(j) which is the relationship score between the test image and every single image in the comparison set. A set of P (x)(P1(x), P2(x), ..., PK(x)) can be obtained based on Eq. (7)