search

idealo / image-quality-assessment

Convolutional Neural Networks to predict the aesthetic and technical quality of images.
https://idealo.github.io/image-quality-assessment/
Apache License 2.0
2.09k stars 447 forks source link

Can the model reach the loss value and accuracy given in the paper? #94

Open Adenialzz opened 3 years ago

Adenialzz commented 3 years ago

Can the model reach the values given in the paper with backbone network is MobileNet, VGG16, Inception-V2 respectively. And according to my experiments, It can not. So I wonder if there something wrong in my implementation. It will be appreciated if you can share your experiment result with me.

Asomey commented 2 years ago

hello bro, i want to ask a question that how can l evaluate the network‘s performance. Since i saw many evaluation standards given by authors like EMD,LCC,SRCC. How can i calculate these evaluation standards. Thanks bro.

Adenialzz commented 2 years ago

hi, bro. The scipy library already provides a method to calculate the correlation coefficient: from scipy.stats import pearsonr from scipy.stats import spearmanr As for emd, you can check the src/utils/losses.py file in this repo.

gohna commented 2 years ago

Hello, I would like to ask whether there is a big difference between your Mobilenet and VGG, because I need pre-trained NimA for other tasks, but the result I directly predicted by using this pre-trained model is quite different from that in the paper, I want to know whether it is the problem of Backbone or the startup parameters of Nima