I have two identical autoencoders with the same number of layers and parameters. I would like to find the similarity between two images from different domains such as an image captured from the camera and another one its sketch.
AE1(Autoencoder) produces the feature vector of camera images and the AE2 feature vector of the sketch-based image.
In the later stages, both feature vectors will be put in Euclidean or another metric network which returns the similarity score [0, 1].
In this way, both AEs learn different weights.
On the other hand, I can use a single AE network with 2 instances. In this way, the weights will be shared.
I heard the term „constraint the network against each other“. Does it refers to case 1, which learn the different weights or case 2, which refers to learn similar weights. Or does it mean something else?