A fundamental question!

[arianaa30]

I just downloaded your latest release package. And simply running evaluate_mobilenet.py on these two images (replaced img_path = 'images/girl1.jpg'):

While there is a huge difference visually, the good photo's score is even lower! Am I missing any configurations?!

Evaluating :  images/girl1.jpg
NIMA Score : 4.269 +- (2.027)
Evaluating :  images/ugly.jpg
NIMA Score : 4.317 +- (1.792)

[mr-Mojo]

I have the same problem, the following images receive ratings that do not make any sense. The bright, beautiful image is score with 6.485, and the dark version which is clearly inferior receives a 6.669. Any thoughts on why this might be?

[guantinglin]

Hi, mr-Mojo and arianaa30 Have you figure out where the problem is?

[mr-Mojo]

Hi @guantinglin, I think that NIMA might be insensitive to lightning and brightness change in images. I confirmed this observation by testing with different implementations from GitHub and my own research. The difference in prediction score is probably due to changes in contrast etc., that come with brightness changes.

[arianaa30]

NIMA also provides a technical score for things like brightness, contrast, and PSNR-like quality of images. That's a different one from Aescetics.

[h3clikejava]

@arianaa30 Have u figure out?

[arianaa30]

No I didn't dig more into this after that. What I did figure out was it gives higher rates to faces and natural images in general. Maybe not really darkness/brightness. I didn't test the latter.

On Wed, Dec 16, 2020, 9:43 PM H3c [email protected] wrote:

@arianaa30 https://github.com/arianaa30 Have u figure out?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/titu1994/neural-image-assessment/issues/59#issuecomment-747185311, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABKBFTIN5TTRRELX63NZDRDSVF457ANCNFSM4IMCN4XQ .