search

uber-research / jpeg2dct

Other
258 stars 45 forks source link

Resizing DCT #16

Open saitarslanboun opened 3 years ago

saitarslanboun commented 3 years ago

Hi!

Thanks for such a cool work!

I use your library for my Pytorch based model. I use fixed-size images, transform them to DCT format using your library, and stream them using data iterator.

Now, I would like to change my data iterator. Instead of using fixed image inputs, I would like to dynamically change the image size at every iteration, not epoch.

Is applying bilinear interpolation directly on DCT output equivalent to applying bilinear interpolation to images, and afterwards applying DCT?

If there is no big difference, I would like to go for the 1st one, as doing it is very simple. If I have to go for 2nd, I have to spend more time.

Thanks,

DLumi commented 7 months ago

Hi, @saitarslanboun

Not sure you need this, like, 4 years later, but maybe this would help to someone later on. Generally speaking, any upsampling operations you do in frequency domain (DCT coefficients specifically) do not translate to ones done in spatial domain. So if you need pixel manipulation of any kind, you i-DCT, do your manupulations, and then DCT again.