A Random Forest-based No-Reference Quality Metric for UGC Videos

Abstract

The images and videos randomly captured by people for sharing in social media or their own use are commonly termed as the user generated content. With the growing popularity of various social media and streaming platforms along with the availability of low-cost portable devices, the amount of such user generated content is exponentially increasing. Visual quality assessment of such user-generated contents is necessary for various purpose such as estimating how the distortions induced during media transmission effects the visual quality of the non-professionally captured image or video, the social media platforms to estimate quality of a media before it gets posted, assessing quality to evaluate performance of a handheld camera or mobile phone, etc. This is a very challenging task due to the fact that the user generated content significantly suffers from multiple artifacts and distortions during both the capturing and transmission pipeline stage that eventually hinder the visual quality. This work mostly deals with the artifacts induced during video capturing stage, and subsequently, use them for estimating visual quality. A random forest-based no-reference video quality assessment metric is proposed for user generated content videos. The proposed approach is divided into two steps. Firstly, the encoding and content-based features are extracted at the frame level. Secondly, an ensemble-based prediction model is employed to exploit the extracted frame-level features for estimating the visual quality score for each frame. Finally, max pooling is performed to estimate the final video level quality score. We also study various score predictors to eventually suggest the best performing ensemble-learning method for the proposed model. Experiments are performed on the benchmark ICME grand challenge dataset of user generated content videos. The model is compared with several state-of-the-art user generated content video quality metrics. The observed results indicate that the proposed no-reference model outperforms the existing approaches for quality estimation of user generated content videos.