vztu / videval

[IEEE TIP'2021] "UGC-VQA: Benchmarking Blind Video Quality Assessment for User Generated Content", Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, Alan C. Bovik

Hi there, I am Zhengzhong Tu 👋

我正在招生，详情请见个人主页.

I'm an Assistant Professor of Computer Science at Texas A&M University. Previously, I was an AI researcher at Google Research.

I graduated from UT Austin ECE, where I worked with Professor Alan Bovik in the LIVE Lab.

📫 Contact me by:

• Website
• 知乎:「假熊猫」
• Twitter
• Linkedin
• Scholar

videval's Issues

sklearn error

1. scikit-learn 0.20.3 cannot be installed in python3.8 enviroment
2. when install scikit-learn latest version, when execute program it generate error about sklearn and joblib

please tell me which env you use! thanks you!

version of sklearn

please, which version of sklearn is used in this model?

Results on the H265 codec videos

Can run in Octave?

Thanks for your great works! Can this model run in Octave not in Matlab?

About the feature selection

Hi，

I wonder if we need first normalize the features before using the sequential feature selection. As the regression model takes SVR, which requires the normalized input.

I checked the source code. The SFS will not automatically call the minmax norm within it. I notice in your script the SFS input is not normalized. Why ?

Please give a “requirements.txt” file in the python section

Please give a “requirements.txt” in the python section，the versions of sklearn and numpy have been incorrect，thank you.

New dataset

Thank you very much for your great work, it is very helpful, I want to test the model on other datasets can you please guide me how to to train it on my dataset please ?, thanks in advance!

Live-VQC dataset

Thanks for your great work. It helps me so much! I tried to download the Live-VQC dataset, but failed to do so every time. Could you please provide me with another way to download the live-VQC dataset? Thank you very much.

Two questions about the codes.

1. Maybe there is a line 'prev_YUV_frame = imresize(prev_YUV_frame, ratio);' lost in VIDEVAL/include/calc_VIDEVAL_feats_light.m line64-66? I found this problem when run this code on LIVE-VQC dataset.
2. Whether there is a mistake in VIDEVAL/demo_eval_BVQA_feats_one_dataset.py line 167 function curvefit? I think the two params 'y_param_valid_pred, y_param_valid' should be 'y_param_train_pred, y_param_train' instead compared with line 217-226.

Will you release a python-version?

Thanks for the sharing. Do you have the plan to release a python version of this code ?

Could I train and test this model on my own database?

pooling strategy

I wonder when we study "Effects of Temporal Pooling", if we pool the features or the scores ?
If scores, how do we get the predicted score of every frame?
If features, how do we pool the features, if do to every dimension of features, eg. 36 dimension in Brisque？

Library not loaded: @loader_path/libmex.dylib

When I run demo_compute_VIDEVAL_feats.m, there is an error:
dlopen(/Users/XXX/VIDEVAL_release/include/matlabPyrTools/corrDn.mexmaci64,
6): Library not loaded: @loader_path/libmex.dylib

Can you give me some help?

light version

A question for the grid search

Hi Zhengzhong,

I noticed in your evaluate script (demo_eval_BVQA_feats_all_combined.py), a grid search is used to obtain a best [gamma, c] for each iteration.

This may bring a problem: for each iteration (the 8-2 division) the best gamma c will be different.

If we have 100 params pairs, and 100 iterations. Then we should have a 100x100 searching matrix, naming K. Usually I take the median for each row (i.e. each iteration with a fixed param pair), then the max value to select the best param pair.

It takes the np.max(np.median(K, 1)) as the final result, here the np denotes numpy.

The current script essentially takes the np.median(np.max(K, 0)).

I don't think these two ways will make the same results ---- the latter one may offer a better score but it can not tell us what's the best params in the whole dataset.

Please correct me if I misunderstand anything....