TECHNICAL
Changing video consumption patterns Why classical video metrics are no longer enough More recently, this challenge is not so easy to assess using traditional engineering metrics. PSNR or other metrics may still technically be useful, but they don’t always correlate well with human perception. Even if the signal is mathematically cleaner, that doesn’t make it subjectively a better viewing experience when motion handling and/or compression artefacts or sharpening characteristics are visually distracting. That is the reason behind gradual industry movement to perceptual quality-based models. For instance, Netflix has proposed the VMAF framework, which utilises machine learning along with human vision modelling to give a more accurate approximation of observed video quality. In a similar vein, standards bodies such as the International Telecommunication Union (ITU) have approached subjective QoE evaluation methods closer to real-life viewing behaviour through the issuance of ITU-T recommendations P.910 and P.1203 [6]. The implications are significant. Video optimisation is not just a simple exercise in signal fidelity. It has become a simultaneous optimisation problem involving user perception, network economics, device heterogeneity and infrastructure economy.
the contribution content-adaptive optimisation can make as part of a state-of-the-art video delivery platform. The research evaluated the performance of VisualOn’s optimisation layer across multiple transcoding scenarios, such as software H.264 and H.265 encoding, NVIDIA NVENC graphics processing unit (GPU) acceleration, NETINT ASIC-based transcoding and Intel QuickSync hardware acceleration. The flow in the QoE does take a path where quality control is difficult as it moves through the networks. As in the diagram below, after the ABR ladder is the last area where there is a chance to adapt the quality before releasing this. The Content Delivery Network (CDN) and the broadband network transmit the quality they receive, hopefully with little added erroneous signals.
Source Video ▼ Encoder ▼ VisualOn Optimizer ▼ ABR Ladder ▼ CDN ▼ Broadband Network ▼ Customer Device and QoE Feedback.
The Limits of Conventional Transcoding
The significance of the findings is not only intrinsic to perceptual video quality improvement, but also to where that improvement occurs. The biggest gains were seen in the lower and middle tiers of the adaptive bitrate (ABR) ladder, across nearly all platforms tested. These are the resolutions most typically served when viewers have either restricted bandwidth on their networks, busy Wi-Fi, or changing mobile connectivity. Enhancements at these resolutions, consequently, have an enhanced impact on perceived Quality of Experience (QoE) for end users. H.265 software encoding showed only small improvements in the highest resolutions, which reflects an already-mature process for H.265, at least compared to older codecs. Still, the advantages really did not start to show until lower resolutions. For instance, the 1080p profile changed from a 90.56 VMAF to a VMAF of 92.95 and the 720p profile from an 84.45 VMAF up to a VMAF score of 88.57, the VMAF rose to 82.05 versus the baseline 76.02 VMAF at 540p (an increase of over JND, or “Just Noticeable Difference”). These results show that simple tools can be deployed to significantly enhance perceptual quality at constant bitrate, allowing
In the past, operators and streaming platforms have struggled to find a suitable compromise between software and hardware transcoding methodologies. While software encoding environments offer great control and flexibility with quality management, they consume a lot of compute resources and energy. While hardware acceleration platforms offer advantages of scale, density and efficiency, they have traditionally come with assumptions of lower quality performance — especially at low bitrates. With streaming volumes ever-increasing, that compromise is no longer acceptable. Operators have been forced to adopt transcoding architectures able to deliver scale and perceptual quality at the same time. So the debate is shifting away from simple “software versus hardware” arguments to a much more fundamental question: how smart can we make the coding pipeline in and of itself?
VisualOn and Cires21 comparative analysis
The joint comparative analysis conducted by VisualOn and Cires21 offers useful guidance on
109
Made with FlippingBook - Online magazine maker