In QoMEX'18, we have proposed a distortion specific No-Reference Video Quality Metric (NR-VQM) for detecting packet loss artifacts. This is implementation of a distortion specific No-Reference Video Quality Metric (NR-VQM) for detecting visibility of packet loss artifacts. Unlike most of the other NR-VQMs known in the prior art, the proposed scheme operates at the frame level, not the sequence level, and therefore it can be used to estimate the position of the impacted frames, rather than assessing the sequence level quality only. The metric is based on hand-crafted features and conventional learning-based regression. Since the proposed metric uses distortion-specific features, it is computationally less complex than most general purpose video and image quality metrics.
The following files are included:
Use this Matlab function to compute FR and NR features for a video sequence and write them in a CSV file for further processing.
This Matlab script shows an example how to use compute_features.m. For using the script, EPFL-PoliMi 4CIF video sequences need to be downloaded and decoded. You can download the database provided by the authors from http://vqa.como.polimi.it/.
This Python script shows an example how to train and validate a regression model to predict frame and sequence level quality scores, using the features computed with EPFL_PoliMi_4CIF_example.m. The script implements "leave-one-out" validation, without using the additional training contents.
More details about the method are available in the following publication:
J. Korhonen, “Learning-based prediction of packet loss artifact visibility in networked video,” IEEE International Conference on Quality of Multimedia Experience (QoMEX’18), Sardinia, Italy, May 2018.