A Blind Quality Metric based on Hybrid Training for HDR VQA
University of Bristol
About
This work describes a new blind video quality assessment (BVQA) method based on a hybrid training methodology, BVQM-HT, which was submitted to the WACV 2023 HDR Video Quality Measurement Grand Challenge. This quality metric was first trained on a large amount of training material from a diverse video database, BVI-DVC, and the training data were labelled using an existing perceptual quality metric, VMAF. An additional shallow CNN is then employed for temporal pooling, optimised using a small HDR subjective video database, provided by the Challenge. This no-reference quality assessment method was then evaluated on the Challenge private test set, and achieved superior correlation performance with perceptual quality over other submitted BVQA methods, with a SROCC value of 0.749.
Source code
Source code from github will be avaliable very soon.Model
Results
The evaluation performance of all BVQA methods in the WACV 2023 HDR Video Quality Measurement Grand Challenge.
