I have got q on imagenet stats. If i understand correctly imagnet stats is based on statistics of images used for training resnet using Imagenet dataset.
Original set of images could differ from images that we end up using for training Resnet model
Why do we use the Imagenet stats for nomalizing the new images lets say eg. Satellite images in Planet Amazon multilable classification ,which could be vastly different from images of Imagenet dataset?
We could have also used mean & std of satellite images…
Can some one please explain
i have the similar doubt buddy,please let me know if it got clear to you
i used different stats in human protein competition… my results were not awful.
So if we think there can be vast difference between images of imagenet and non imagenet one can still use the stats of different images
Someone please correct me if I am wrong, but I think we use imagenet because we can benefit from base layers information, such as shapes, corners, shades, textures, etc. For more top layers, like dog’ eyes or noses or something, we should unfreeze and retrain model on our own layers, I think? (but I could be wrong, I am a newbie)
I think normalising brings the numbers into the realm of what the model has already learnt to understand. So it just helps it converge faster