The same image data and resize to 6464 and 128128, but the latter of result is not convergence

9 views

Skip to first unread message

unread,

Nov 8, 2016, 4:40:02 AM11/8/16

to Caffe Users

I used webface image data to train and test caffe.

First, detected face.

Second, resized all image to 64*64 and 128*128.

The result is very different. The test accuracy of size of 64*64 that is 0.83 and convergence, but the size of 128*128 cann't converge.

This two size used training net and solver.prototxt is almost same, only the batch_size of train and test are differnet.

The batch_size of train and test that the size of 64*64 are 256 and 128, then the batch_size of train and test that the size of 128*128 are both 32.

The same image data and resize to 64*64 and 128*128, but the latter of result is not convergence