1

I'm new to Caffe framework and I want to create a siamese network.

My dataset has multiple images with eyes: x_left.jpg and x_right.jpg: left eye and right eye for person x and the train sample looks like:

__image__, __level__
15_left  ,      1
15_right ,      2
16_left  ,      4
16_right ,      4

Level indicates how advanced is the disease - diabetic retinopathy, so the level can be considered as a category from 0 to 5.

The two inputs are going to be two images, i.e. 15_left.jpg and 16_right.jpg. After training the siamese network, I want to use it as an extractor feature, i.e. Give it a picture and keep features on the penultimate layer.

My question is: how can I create the lmdb files to use siamese network?

Thank you!

Mada
  • 11
  • 2
  • 1
    i am not sure if you know the rules on stackoverflow...if you havent tried something and post output...you won`t get much feedback – Eliethesaiyan May 23 '17 at 02:03
  • @Eliethesaiyan got a point. your question is not clear. Let's try and improve your question: (1) [format](https://stackoverflow.com/help/formatting) your text. (2) Describe in a *little* bit more detail what you are trying to learn with your net, what are the two inputs going to be? what do the labels mean? Please edit your question to reflect these changes. – Shai May 23 '17 at 05:23
  • Thank you for your answers. I edited the question. – Mada May 23 '17 at 08:02
  • why do you need a Siamese network here? do you want the same net for both left and right eyes? do you want to compare the `__level__` of the left and right eyes? why not simply classify the eye images according to their respective `__level__`? – Shai May 23 '17 at 08:16
  • yes, I want to use the same net for left and right eyes. This is an experiment for school project. I think an idea is to compute the similarity: label 1 if the images are from same person and label 0 - images are from different persons. I can do this using just the image name. but I don't not how to create the lmdb and what layers to use for data/input. – Mada May 23 '17 at 09:44

0 Answers0