small training set along with its sampling strategy

[liangjianfans]

Hi Marvin,

Very nice work. I am trying to adopt your code to my data set, which has only 100 image pairs. Each image has size of 800*800. The mask image has only two classes: background and foreground. Do you think this image set is too small?

I am trying to generate small patches (256*256) from the original image. Across the image set, the foreground area takes about 25% of the whole image area, as shown in the attached figure. Are there any recommended strategy for generating these small patches? Moreover, during the training process, do we need to enforce all the patches sampled from the same image to be put into a single patch?

Thanks, Liangjian