Can anyone help me understand per-pixel loss versus content loss?

Part 2 (2017)
1

I couldn’t find the definition of per-pixel loss and am struggling to understand the difference between per-pixel loss and content(perceptual) loss. I understand that

  1. both loss functions measure how similar 2 images are alike.
  2. for content loss you can combine multiple layers so that’s an advantage.

In terms of measuring the loss, are they both measuring the per-pixel differences? if not, how are they measuring things differently? thank you!

2

per-pixel loss = measuring the actual differences between pixels in the images. i.e. it’s very naive and fancy sounding for a basic concept.

3

Thanks @kelvin!

Can you verify if the following description is correct? If not, can you point out the errors for me?

for “per-pixel loss”, we are summing all the absolute error between all pixels.
for content loss, we are summing all the squared errors between all pixels and take the mean.

4

The loss function is interchangeable conceptually. i.e. you can use MAE/MSE for either/both per-pixel and content loss.

5

I think I follow that, so what’s the difference per-pixel loss and content loss (if it’s not about the loss functions they use)?

6

I misread your earlier comment.

per-pixel = measuring differences between pixels.
perceptual = measuring differences between a specific layer’s filter activations in a convnet.

3 Likes
7

OHHHHHHhhhhhhhhh
Thank you! You just saved me a massive headache!!!

1 Like