results of different experiments?

[martinbel]

Hi,

I'm curious about how other experiments performed, it would be great to add these results to the readme if possible. For example, have you tried using the entire image as input instead of small sized batches? This is basically the idea of the U-Net and worked well in other tasks. Thanks again for open sourcing this!

[lantiga]

Hi Martin, we're currently running the model on the STARE dataset. We'll be publishing results soon.

As for the entire image, we would have gladly tried it, but the DRIVE dataset only consists of 20 images (STARE is not a whole lot more). We'd need a much larger dataset of annotated cases to be successful I believe.

We have developed a ladder network / U-net hybrid internally, which (in theory) helps with semi-supervised segmentation tasks. We could take advantage of an un-annotated dataset to build up a robust whole-image net, we'll eventually go this route.

Stay tuned for the next batch of results and feel free to contribute additional experiments.

[GlastonburyC]

Can you detail the size of the training dataset used in the original U-net paper? I believe it's very small.

[fschi]

In the original U-Net paper they used 30 Images with 512x512pixels