Knowledge Guided Semi-Supervised Learning for Quality Assessment of User Generated Videos

The deep learning based approaches particularly require training on large amount of labelled data, which is Perceptual quality assessment of user generated content cumbersome and expensive to acquire. This leads to us to (UGC) videos is challenging due to the requirement of the question of how we can design NR VQA models which large scale human annotated videos for training. In this can be trained with very limited labelled training data, yet work, we address this challenge by first designing a selfsupervised achieve excellent generalisation performance on multiple Spatio-Temporal Visual Quality Representation datasets in terms of correlation with human perception. Learning (ST-VQRL) framework to generate robust quality Our focus in this work is on designing semi-supervised aware features for videos. Then, we propose a dual-model NR VQA method with limited labelled along with unlabelled based Semi Supervised Learning (SSL) method specifically data. Since UGC videos have diverse quality characteristics, designed for the Video Quality Assessment (SSL-VQA) task, we believe that pretraining a robust video quality through a novel knowledge transfer of quality predictions feature backbone is extremely important to transfer knowledge between the two models. Our SSL-VQA method uses the during semi-supervised learning. With this motivation, ST-VQRL backbone to produce robust performances across we approach the problem using a combination of various VQA datasets including cross-database settings, contrastive self-supervised pretraining followed by semisupervised despite being learned with limited human annotated videos.