how to use the ground truth image as the label in caffe

1,150 views

Skip to first unread message

dearl...@gmail.com

unread,

May 12, 2016, 2:46:17 AM5/12/16

to Caffe Users

Hi everyone,

I am trying too train a network about segment, and the label is the ground truth images the same size as the input images,

i do not know how to use the ground truth image as the matrix label in caffe ,could anyone give me some advise or examples

about the issue.

thank you.

Clint Sebastian

unread,

May 13, 2016, 3:28:52 AM5/13/16

to Caffe Users

Hi,

You can create two database layers (one for the training data and other for the groundtruth) and then calculate the euclidean loss.

You can check it here. https://github.com/BVLC/caffe/issues/1013

The last layer definition in the prototxt file can be found here. https://github.com/BVLC/caffe/issues/2818

Jan

unread,

May 13, 2016, 4:44:31 AM5/13/16

to Caffe Users

Clint is right. Moreover, here you find everything you need to know about FCNs. Note, that opposed to what the notice there is saying, you don't need to use the longjohn/future branch any longer, everything has been integrated into caffe master by now.