How to train model using part of image


Hello everyone,
I have set of images to train a model in MXNET. However, the images are such that real subject in image is at center and covers less than 40% of image. Rest portion is just blank white background.

Based on this scenario, can I specify just a part of image to build MXNET model?

Will it help if I crop the image to cover only the subject in it?
My validation images in real time will also be in same format (Subject placed in center and cover less than 40% of plain white background). So if I crop my image while training model, do I need to do similar cropping during realtime validation as well?

I am OK for any other alternatives as well.

Thanks in advance.


@abalki This is a good question. Could you tag this as Discussion? That might receive more general audience attention.


Anyone came across this scenario before?
Since subject in training images does not occupy major portion, final model can classify major differences easily but not minor defects.