In some deep learning papers i read about multiscale inputs, so i wanted to read about scale of an image. What i got to know is that fundamentally scale is related to the distance of the object being captured from the camera. source from where i read

But to produce images at different scale i see that one smoothes it with gaussian filter, and varying its parameters we get image at different scales. This doesn't intuitively makes sense to me because objects in the image still remain at same size, they just get even more blurred. This way of achieving it seems to not be capturing the fundamental motivation behind image scale.

Can someone please clarify ?


1 Answer 1


Multiscale inputs in the context of deep learning usually refer to using several copies of the same image with reduced resolution.

For example, the default image may be 400x400 pixels large. It is then rescaled to 200x200 pixels, 100x100 pixels and 50x50 pixels. Now you have the same image represented on four "different scales" because the size of one pixel in each version capture differently large part of the physical world. The blurring is performed to better preserve the information lost during the downscaling of the images.

This is also called image pyramid: enter image description here

  • $\begingroup$ This image pyramid is very much related to the scale that i am asking about cs.uu.nl/docs/vakken/ibv/reader/chapter9.pdf $\endgroup$ Jan 9, 2019 at 9:27
  • $\begingroup$ Is there still something unclear? $\endgroup$ Jan 9, 2019 at 13:28
  • $\begingroup$ no you have given me whole thesis worth of knowledge , what can be unclear then $\endgroup$ Jan 9, 2019 at 14:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.