Custom Dataset for Siamese CNN

vision
1

Hi all,
I’m building a particular kind of dataset for a Siamese neural network.
I was inspired by this topic: Dataloader for a Siamese Model with ConcatDataset, but my implementation is slightly different since I need to take the 2 images, passed to the net, from 2 different folders and I need that the first and the second one images are always taken from the respective dataset.
So I tried the implementation linked but it doesn’t fit my problem because the Concat function don’t respect the constraint discussed.
Thus, my question is: is there a way to tell the Concat function to respect these constraint?

P.s. The 2 dataset/folders are of different size so I think that the one with less images must recycle some of its images to be combined with others of the second bigger dataset

2

Anyone got some tips for this task?

3

Since your use case seems to be a bit more complicated I would recommend to implement a custom Dataset as explained in this tutorial.
This would allow you to implement the loading logic (and correspondence of the image folders) manually.