Temporal attention networks for no-reference video quality assessment

Using attention to extend image quality assessment algorithms for the task of video quality assessment.

Two-stage learning

Attention is used over features extracted by SOTA no-reference IQA algorithms in attention.py.

End-to-end learning

Paq-2-Piq (an IQA model that uses deep-learning) is paired with an RNN and attention (CNN+RNN framework) to learn VQA scores end-to-end in vqa.py.

ann.py trains feature vectors that are averaged across frames. attention.py trains an attention network to weigh frame features before training. vqa.py performs end-to-end training with attention and Paq-2-Piq.

Please cite this repository if you use this in your work.