Is it better to always expand our image size?

In the first week’s videos, the lecturer says:
"training the network with a 32 by 32 image directly will produce very poor results.
So to obtain a better performing model we’ll upscale the images to 224 by 224 before sending these as inputs into the resnet 50 layer object "

does he mean in general, training on larger image produce a better result? or he just means for this particular example, 224*224 is better since it has been pre-trained on images of this size?

Hi, @majid_mmd!

Upsampling images is usually not a good idea since you are just wasting memory and computation adjusting more weights due to replicated copies of pixels, so technically you cannot extract more information than with 32x32 input. Nonetheless, I’m sure the 224x224 model is way more optimized in terms of feature extraction than the 32x32, so this and the pretrained weights may be the reasons of the improvement.

1 Like