Subjective and objective quality assessment for advanced videos



Journal Title

Journal ISSN

Volume Title



The surge of video streaming services, particularly for high motion content such as sporting events, necessitates advanced techniques to maintain video quality, facing challenges such as capture artifacts and distortions during coding and transmission. The advent of High Dynamic Range (HDR) content, offering a broader and more accurate representation of brightness and color, poses additional complexities due to increased data volume. The critical need for robust Video Quality Assessment (VQA) models arises from these challenges. To meet this need, we conducted three substantial subjective quality studies and constructed corresponding databases. The Laboratory for Image and Video Engineering (LIVE) Livestream Database comprises 315 videos of 45 source sequences from 33 original contents impaired by six types of distortions. This database facilitated the gathering of over 12,000 human opinions from 40 subjects. The LIVE HDR Database, the first of its kind dedicated to HDR10 videos, includes 310 videos from 31 distinct source sequences, processed with ten different compression and resolution combinations. This resource was instrumental in amassing over 20,000 human quality judgments under two different illumination conditions. An additional LIVE HDR AQ was developed with 400 videos from 40 unique source sequences. These videos were processed using varied compression, resolution combinations, and AQ-mode settings, to study the effects of adaptive quantization (AQ) and rate-distortion optimization techniques on HDR video perceptual quality. Building on these invaluable databases, we developed two innovative objective quality models: HDRMAX and HDRGREED. HDRMAX, a pioneering framework designed to create HDR quality-sensitive features, augments the widely-deployed Video Multimethod Assessment Fusion (VMAF) model, yielding significantly improved performance on both HDR and SDR videos. HDRGREED, a novel model leveraging localized histogram equalization and Difference of Gaussian filters, employs the Generalized Gaussian Distribution to model the bandpass responses and measure the entropy variations between reference and distorted videos. This model is particularly sensitive to banding and blocking artifacts introduced by inappropriate AQ settings. In conclusion, the comprehensive subjective quality studies and databases, along with the state-of-the-art objective quality models, HDRMAX and HDRGREED, significantly contribute to the advancement of future VQA models. These tools cater specifically to challenges posed by live streaming and HDR content, providing critical resources for the development, testing, and comparison of future VQA models. These databases, publicly available for research purposes, and the innovative models offer valuable insights to improve and control the perceptual quality of streamed videos.


LCSH Subject Headings