GoogleNet / AlexNet (caffenet): how the crop size 224 vs 227 is determined?

MW

unread,

Mar 19, 2016, 5:17:48 AM3/19/16

to Caffe Users

Hi,

I saw in https://github.com/BVLC/caffe/blob/master/models/bvlc_googlenet/train_val.prototxt

crop_size: 224

and in https://github.com/BVLC/caffe/blob/master/models/bvlc_alexnet/train_val.prototxt

crop_size: 227

and in https://github.com/BVLC/caffe/blob/master/models/bvlc_reference_caffenet/train_val.prototxt

crop_size: 227

I'm just wondering how these "magic" number is determined? why not just use the original 256x256 image size?

What's the reason for doing these? Any one can give me some insight?

Thanks!

Christos Apostolopoulos

unread,

Mar 19, 2016, 7:23:36 PM3/19/16

to Caffe Users

It's for data augmentation. Basically if you have a 256x256 image you can crop a number of 227x227 patches and thus increase your training size without actually adding any real data. You can also mirror your images for the same effect (command mirror:1)

Oscar Beijbom

unread,

Mar 21, 2016, 4:17:57 AM3/21/16

to Caffe Users

224 vs 227 is an arbitrary feature of the network architecture.

Reply all

Reply to author

Forward