I'm approaching at Siamese Networks in order to use them for Image Similarity. I found that many people use famous models like VGG or ResNet to build the vectors that will go on the distance layer in order to be compared with the other vector image and then they use contrastive loss. Then I found that other people added some other layers after the distance one in order to get in output with a sigmoid a similarity label (0,1 or discretized).
Is this final processing useful? These networks are born in order to work with the distance vector right?
So, I'm trying to understand the main differences between the two approaches, if any.