0

I have two images/patches (the larger one includes the smaller one and some context information around it) of different sizes (I dont wont to create a new image by just concatenating them).

I want to create a CNN as shown in the figure.

How can I create it using Caffe prototxt file?

Can I directly use Caffe to train it?

Both CNNs can share the same weights.

enter image description here

user570593
  • 3,420
  • 12
  • 56
  • 91
  • Are the CNN outputs of the same shape? This would be necessary to feed them into the same FC layer, right? – Prune Jul 06 '16 at 19:05
  • Also, do you *need* the two CNNs to share the same weights? If so, at what point do they achieve the same shape for the processing? That would be the point of joining the data streams. – Prune Jul 06 '16 at 19:06
  • It is not necessary to share the same weights. The two CNN architectures could be different. The shapes also could be different. Yes. they have to connected to the same FC layer. – user570593 Jul 06 '16 at 20:07
  • Then how do you envision tying them into the *same* FC layer? At what point do you converge the shape dimensions? – Prune Jul 06 '16 at 20:09
  • Dont understand. Each CNN can have its own FC. These two FC can be connected to the next FC layer. – user570593 Jul 06 '16 at 21:56
  • Okay -- perhaps what you see as a problem, I simply see as a straightforward declaration. What stops you from simply giving both CNN outputs as **top** to the FC layer, just as in any other Siamese NN? – Prune Jul 06 '16 at 22:25
  • I am not sure how to prepare the inputs as I have to input a pair of images (img1, img2). – user570593 Jul 07 '16 at 10:30
  • Also I dont know how to use the pretrained weights if I have two CNN as shown in the figure. – user570593 Jul 07 '16 at 11:09
  • @Prune Thanks for the suggestions with Siamese NN. But I am not sure how to initialize the two networks with a pretrained model. – user570593 Jul 07 '16 at 15:12
  • Let us [continue this discussion in chat](http://chat.stackoverflow.com/rooms/116719/discussion-between-prune-and-user570593). – Prune Jul 07 '16 at 17:22

0 Answers0