Learning to Compare Image Patches via Convolutional Neural Networks --- Reading Summary
Learning to Compare Image Patches via Convolutional Neural Networks --- Reading Summary
2017.03.08
Target: this paper attempt to learn a geneal similarity function for comparing image patches from image data directly.
There are several ways in which patch pairs can be processed by the network and how the information sharing can take place in this case. This paper studied 3 types about the comparion network:
1. 2-channel 2. Siamese 3. Pseu-siamese Network
1. Siamese Network :
This is a chassical network which first proposed by Lecun. This network has two networks which denote two inputs (the compared image pairs). Each network has its own convolution layer, ReLU and max-pooling layer. It is also worthy to notice that: the two networks are share same weights.
2. Pseudo-siamese Network :
the same definition as siamese network, but the two branches do not share weights. This is the most difference between siamese and pseudo-siamese network.
3. 2-channel network :
Just combine two input patches 1 and 2 together, and input it into normal convolutional network. The output of the network is 1 value. This kind of network has greater flexibnility and fast to train. But, it is expensive when testing, because it need all combinations of patches.
