The predicted scores are not correct

[sathishkumark27]

Hi, Thanks for sharing the pre-trained models.
I am using the init model and when I try to predict the scores for below images the scores are not correct
image -1 :

Scores : {'fc11_score': array([[0.7237532]], dtype=float32), 'fc9_MotionBlur': array([[-0.00683291]], dtype=float32), 'fc9_Repetition': array([[0.1499309]], dtype=float32), 'fc9_ColorHarmony': array([[0.41779417]], dtype=float32), 'fc9_Light': array([[0.23401856]], dtype=float32), 'fc9_Content': array([[0.44104707]], dtype=float32), 'fc9_Object': array([[0.15831222]], dtype=float32), 'fc9_BalancingElement': array([[0.08793019]], dtype=float32), 'fc9_Symmetry': array([[0.05824818]], dtype=float32), 'fc9_RuleOfThirds': array([[0.2551845]], dtype=float32), 'fc9_DoF': array([[0.14344457]], dtype=float32), 'fc9_VividColor': array([[0.20968038]], dtype=float32)} ./frame3467.png 0.7237532

Image -2

Scores : {'fc11_score': array([[0.66474986]], dtype=float32), 'fc9_MotionBlur': array([[-0.00710997]], dtype=float32), 'fc9_Repetition': array([[0.11318572]], dtype=float32), 'fc9_ColorHarmony': array([[0.3682373]], dtype=float32), 'fc9_Light': array([[0.19498739]], dtype=float32), 'fc9_Content': array([[0.52042776]], dtype=float32), 'fc9_Object': array([[0.13005672]], dtype=float32), 'fc9_BalancingElement': array([[0.04698427]], dtype=float32), 'fc9_Symmetry': array([[0.13396302]], dtype=float32), 'fc9_RuleOfThirds': array([[0.23525675]], dtype=float32), 'fc9_DoF': array([[0.17191094]], dtype=float32), 'fc9_VividColor': array([[0.11277764]], dtype=float32)} ./frame2742.png 0.66474986

By looking at those 2 images it is evident that Image -1 is better than Image-2 but the predictions are suggesting otherwise, can you please help to understand the issue?

[aimerykong]

Hi, thanks for your interest and trying this! The failure results are not too surprising to me. My explanation is that, the model is trained and test on the same image statistics. But here you provide the model with an image which has much artifact, different from the training images. So the model fails easily. This also indicates the limited generalizability of the model. Hope this helps! Shu

…

On Wed, Sep 18, 2019 at 6:10 PM sathishkumark27 ***@***.***> wrote: Hi, Thanks for sharing the pre-trained models. I am using the init model and when I try to predict the scores for below images the scores are not correct *image -1 :* [image: 7237532 14_frame3467] <https://user-images.githubusercontent.com/31283567/65202190-0436b600-da3f-11e9-940f-962887c67828.png> Scores : {'fc11_score': array([[0.7237532]], dtype=float32), 'fc9_MotionBlur': array([[-0.00683291]], dtype=float32), 'fc9_Repetition': array([[0.1499309]], dtype=float32), 'fc9_ColorHarmony': array([[0.41779417]], dtype=float32), 'fc9_Light': array([[0.23401856]], dtype=float32), 'fc9_Content': array([[0.44104707]], dtype=float32), 'fc9_Object': array([[0.15831222]], dtype=float32), 'fc9_BalancingElement': array([[0.08793019]], dtype=float32), 'fc9_Symmetry': array([[0.05824818]], dtype=float32), 'fc9_RuleOfThirds': array([[0.2551845]], dtype=float32), 'fc9_DoF': array([[0.14344457]], dtype=float32), 'fc9_VividColor': array([[0.20968038]], dtype=float32)} ./frame3467.png 0.7237532 *Image -2* [image: 6647498 61_frame2742] <https://user-images.githubusercontent.com/31283567/65202419-32b49100-da3f-11e9-9567-f43aad301d6d.png> Scores : {'fc11_score': array([[0.66474986]], dtype=float32), 'fc9_MotionBlur': array([[-0.00710997]], dtype=float32), 'fc9_Repetition': array([[0.11318572]], dtype=float32), 'fc9_ColorHarmony': array([[0.3682373]], dtype=float32), 'fc9_Light': array([[0.19498739]], dtype=float32), 'fc9_Content': array([[0.52042776]], dtype=float32), 'fc9_Object': array([[0.13005672]], dtype=float32), 'fc9_BalancingElement': array([[0.04698427]], dtype=float32), 'fc9_Symmetry': array([[0.13396302]], dtype=float32), 'fc9_RuleOfThirds': array([[0.23525675]], dtype=float32), 'fc9_DoF': array([[0.17191094]], dtype=float32), 'fc9_VividColor': array([[0.11277764]], dtype=float32)} ../../LLFF/results/horses/aadb_grid/rendered_views/frame2742.png 0.66474986 By looking at those 2 images it is evident that Image -1 is better than Image-2 but the predictions are suggesting otherwise, can you please help to understand the issue? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#24?email_source=notifications&email_token=ABRJSJDFP6TA6IUBNQ46MTLQKLGOPA5CNFSM4IYFHJ3KYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HMJCKAQ>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABRJSJHZ5M3TVS3UQD6EUPTQKLGOPANCNFSM4IYFHJ3A> .

[sathishkumark27]

Okay got it, thank you.