training using pixel wise labeled ground truth

[lei Wang]

Hi guys,

I am trying to do training on pixel-wisely labeled ground truth images. I wonder how should I setup caffe to do this. The data I have consist of around 50 images (each has dimension 2  by 10k by 10k). I also comes with pixel level ground truth images (1 by 10k by 10k). I am trying to use this data to train a CNN. The approach that I am using now is to extract random samples and their corresponding context patches from the images and store the sample patches in hdf5 files for training. The problem is that the number of training samples cannot be very large or the size of the hdf5 files would shoot up quickly. I think a more efficient approach is to load images and generate samples on the fly. Is there a way that I can do this with current caffe implementation? 

[lei Wang]

I forget to mention that the purpose is to do pixel wise prediction (generate prediction of the 10k by 10k).

[Erogol]

Did you come up with a solution ?

[li kai]

did you find a solution？

在 2015年2月14日星期六 UTC+8下午9:08:48，lei Wang写道：