mean subtraction on darknet

asked 2019-09-21 11:48:03 -0500

fogx gravatar image

Does it make sense to use the mean subtraction preprocessing option in blobFromImage before running a darknet DNN? I did not find any information stating that this happens at training time, so it would probably make my results worse if i run it at just inference time right?

Or is this something i should be running on my training dataset (manually) and then enable at inference time? I could not find any clear statements for this while researching online

edit retag flag offensive close merge delete

Comments

1

data with zero mean is better for training

LBerger gravatar imageLBerger ( 2019-09-21 14:17:39 -0500 )edit

Just curious - are you talking about (mean) normalization (values between 0-1) ?

holger gravatar imageholger ( 2019-09-23 16:53:56 -0500 )edit

@holger normalization is a different preprocessing step. Mean Subtraction involves subtracting the mean across every individual feature in the data, and has the geometric interpretation of centering the cloud of data around the origin along every dimension. Here is a description of both

fogx gravatar imagefogx ( 2019-09-24 03:19:58 -0500 )edit

Thank you!

holger gravatar imageholger ( 2019-09-24 09:32:26 -0500 )edit