crop offset size in vgg16_FCN8s

[Joshua Cho]

I am following the FCN for semantic segmentation (describe in https://devblogs.nvidia.com/parallelforall/image-segmentation-using-digits-5/).

For fcn_alexnet, the crop offset size is 18 by calculated backward from last layer. 

But I don't understand how I do get the offset value, 5 and 9 in score_pool4c and score_pool3c layers respectively as below

It would be very thankful if somebody help me to how to calculate the below offset number. 

Thank you

-Joshua

from https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn8s/deploy.prototxt

...

layer {
	name: "score_pool4c"
	type: "Crop"
	bottom: "score_pool4"
	bottom: "upscore2"
	top: "score_pool4c"
	crop_param {
	axis: 2
	offset: 5
	}
	} ...	layer { name: "score_pool3c" type: "Crop" bottom: "score_pool3" bottom: "upscore_pool4" top: "score_pool3c" crop_param { axis: 2 offset: 9 } }

[Evan Shelhamer]

The crop offset is determined by coordinate mapping as part of net specification: https://github.com/BVLC/caffe/blob/master/python/caffe/coord_map.py. I'm sorry to say the process isn't further documented at this time and that you will have to refer to this code for the details of how the coordinates are related across layers. If you search for "coordinate mapping" on the mailing list you will find further discussion in older threads.

My advice is to make use of this net spec helper, and not try to do it manually (except as an aid to understanding!). It's a tedious and error-prone process.

Evan Shelhamer

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users+unsubscribe@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/561d3535-b92d-4711-bf71-c324bdddddd6%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Joshua Cho]

Hello Even,

Thank you for your comments and advices. Can you advice me more details how to make use of net helper for finding the size of crop offset?

You mentioned "net spect helper" is net.py (https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn8s/net.py)?

Thank you

Joshua Cho

Evan Shelhamer

To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.