Batch Normalization introducing noise to segmented outputs

[Clint Sebastian]

Hi,

I am training a network using batch normalization for segmentation.

Apparently, it introduces a lot of noise in my output images although I reach the same amount of loss in a fewer number of iterations.

Has anyone faced the same problem while training?

Any workarounds to fix this?

[Clint Sebastian]

Found the problem!

 

Should turn on the calculated statistics in Batch Normalization layer from training phase during testing.

Something like this during testing phase. 

 

layer { name: "conv1_bn" type: "BatchNorm" bottom: "conv1" top: "conv1"

  batch_norm_param { use_global_stats: true } }