Measuring YouTube’s Perceptual Video Quality – Google AI Blog

Online video sharing platforms, like YouTube, want to know perceptual video high quality (i.e., a consumer’s subjective notion of video high quality) with the intention to higher optimize and enhance consumer expertise. Video high quality evaluation (VQA) makes an attempt to construct a bridge between video indicators and perceptual high quality through the use of goal mathematical fashions to approximate the subjective opinions of customers. Traditional video high quality metrics, like peak signal-to-noise ratio (PSNR) and Video Multi-Method Assessment Fusion (VMAF), are reference-based and give attention to the relative distinction between the goal and reference movies. Such metrics, which work finest on professionally generated content material (e.g., motion pictures), assume the reference video is of pristine high quality and that one can induce the goal video’s absolute high quality from the relative distinction.

However, nearly all of the movies which can be uploaded on YouTube are user-generated content material (UGC), which carry new challenges as a result of their remarkably excessive variability in video content material and unique high quality. Most UGC uploads are non-pristine and the identical quantity of relative distinction might suggest very completely different perceptual high quality impacts. For instance, folks are typically much less delicate to the distortions of poor high quality uploads than of top quality uploads. Thus, reference-based high quality scores develop into inaccurate and inconsistent when used for UGC circumstances. Additionally, regardless of the excessive quantity of UGC, there are at the moment restricted UGC video high quality evaluation (UGC-VQA) datasets with high quality labels. Existing UGC-VQA datasets are both small in dimension (e.g., LIVE-Qualcomm has 208 samples captured from 54 distinctive scenes), in contrast with datasets with hundreds of thousands of samples for classification and recognition (e.g., ImageNet and YouTube-8M), or don’t have sufficient content material variability (sampling with out contemplating content material data, like LIVE-VQC and KoNViD-1k).

In “Rich Features for Perceptual Quality Assessment of UGC Videos“, revealed at CVPR 2021, we describe how we try to unravel the UGC high quality evaluation downside by constructing a Universal Video Quality (UVQ) mannequin that resembles a subjective high quality evaluation. The UVQ mannequin makes use of subnetworks to investigate UGC high quality from high-level semantic data to low-level pixel distortions, and supplies a dependable high quality rating with rationale (leveraging complete and interpretable high quality labels). Moreover, to advance UGC-VQA and compression analysis, we improve the open-sourced YouTube-UGC dataset, which comprises 1.5K consultant UGC samples from hundreds of thousands of UGC movies (distributed underneath the Creative Commons license) on YouTube. The up to date dataset comprises ground-truth labels for each unique movies and corresponding transcoded variations, enabling us to raised perceive the connection between video content material and its perceptual high quality.

Subjective Video Quality Assessment

To perceive perceptual video high quality, we leverage an inside crowd-sourcing platform to gather imply opinion scores (MOS) with a scale of 1–5, the place 1 is the bottom high quality and 5 is the best high quality, for no-reference use circumstances. We acquire ground-truth labels from the YouTube-UGC dataset and categorize UGC components that have an effect on high quality notion into three high-level classes: (1) content material, (2) distortions, and (3) compression. For instance, a video with no significant content material will not obtain a top quality MOS. Also, distortions launched in the course of the video manufacturing part and video compression artifacts launched by third-party platforms, e.g., transcoding or transmission, will degrade the general high quality.

MOS= 2.052	MOS= 4.457

MOS= 1.242	MOS= 4.522

MOS= 2.372	MOS= 4.646

We reveal that the left gaming video within the second row of the determine above has the bottom MOS (1.2), even decrease than the video with no significant content material. A potential clarification is that viewers could have greater video high quality expectations for movies which have a transparent narrative construction, like gaming movies, and the blur artifacts considerably cut back the perceptual high quality of the video.

UVQ Model Framework

A standard technique for evaluating video high quality is to design subtle options, after which map these options to a MOS. However, designing helpful handcrafted options is tough and time-consuming, even for area specialists. Also, probably the most helpful current handcrafted options had been summarized from restricted samples, which can not carry out effectively on broader UGC circumstances. In distinction, machine studying is changing into extra distinguished in UGC-VQA as a result of it may possibly routinely study options from large-scale samples.

A simple strategy is to coach a mannequin from scratch on current UGC high quality datasets. However, this might not be possible as there are restricted high quality UGC datasets. To overcome this limitation, we apply a self-supervised studying step to the UVQ mannequin throughout coaching. This self-supervised step allows us to study complete quality-related options, with out ground-truth MOS, from hundreds of thousands of uncooked movies.

Following the quality-related classes summarized from the subjective VQA, we develop the UVQ mannequin with 4 novel subnetworks. The first three subnetworks, which we name ContentNet, DistortionNet and CompressionInternet, are used to extract high quality options (i.e., content material, distortion and compression), and the fourth subnetwork, known as AggregationNet, maps the extracted options to generate a single high quality rating. ContentNet is skilled in a supervised studying vogue with UGC-specific content material labels which can be generated by the YouTube-8M mannequin. DistortionNet is skilled to detect widespread distortions, e.g., Gaussian blur and white noise of the unique body. CompressionInternet focuses on video compression artifacts, whose coaching information are movies compressed with completely different bitrates. CompressionInternet is skilled utilizing two compressed variants of the identical content material which can be fed into the mannequin to foretell corresponding compression ranges (with a better rating for extra noticeable compression artifacts), with the implicit assumption that the upper bitrate model has a decrease compression degree.

The ContentNet, DistortionNet and CompressionInternet subnetworks are skilled on large-scale samples with out ground-truth high quality scores. Since video decision can also be an vital high quality issue, the resolution-sensitive subnetworks (CompressionInternet and DistortionNet) are patch-based (i.e., every enter body is split into a number of disjointed patches which can be processed individually), which makes it potential to seize all element on native decision with out downscaling. The three subnetworks extract high quality options which can be then concatenated by the fourth subnetwork, AggregationNet, to foretell high quality scores with area ground-truth MOS from YouTube-UGC.

The UVQ coaching framework.

Analyzing Video Quality with UVQ

After constructing the UVQ mannequin, we use it to investigate the video high quality of samples pulled from YouTube-UGC and reveal that its subnetworks can present a single high quality rating together with high-level high quality indicators that may assist us perceive high quality points. For instance, DistortionNet detects a number of visible artifacts, e.g., jitter and lens blur, for the center video beneath, and CompressionInternet detects that the underside video has been closely compressed.

ContentNet assigns content material labels with corresponding chances in parentheses, i.e., automobile (0.58), car (0.42), sports activities automobile (0.32), motorsports (0.18), racing (0.11).

DistortionNet detects and categorizes a number of visible distortions with corresponding chances in parentheses, i.e., jitter (0.112), colour quantization (0.111), lens blur (0.108), denoise (0.107).

CompressionInternet detects a excessive compression degree of 0.892 for the video above.

Additionally, UVQ can present patch-based suggestions to find high quality points. Below, UVQ stories that the standard of the primary patch (patch at time t = 1) is nice with a low compression degree. However, the mannequin identifies heavy compression artifacts within the subsequent patch (patch at time t = 2).

Patch at time t = 1	Patch at time t = 2
Compression degree = 0.000	Compression degree = 0.904

In apply, UVQ can generate a video diagnostic report that features a content material description (e.g., technique online game), distortion evaluation (e.g., the video is blurry or pixelated) and compression degree (e.g., low or excessive compression). Below, UVQ stories that the content material high quality, taking a look at particular person options, is nice, however the compression and distortion high quality is low. When combining all three options, the general high quality is medium-low. We see that these findings are near the rationale summarized by inside consumer specialists, demonstrating that UVQ can motive by means of high quality assessments, whereas offering a single high quality rating.

UVQ diagnostic report. ContentNet (CT): Video recreation, technique online game, World of Warcraft, and so on. DistortionNet (DT): multiplicative noise, Gaussian blur, colour saturation, pixelate, and so on. CompressionInternet (CP): 0.559 (medium-high compression). Predicted high quality rating in [1, 5]: (CT, DT, CP) = (3.901, 3.216, 3.151), (CT+DT+CP) = 3.149 (medium-low high quality).

Conclusion

We current the UVQ mannequin, which generates a report with high quality scores and insights that can be utilized to interpret UGC video perceptual high quality. UVQ learns complete high quality associated options from hundreds of thousands of UGC movies and supplies a constant view of high quality interpretation for each no-reference and reference circumstances. To study extra, learn our paper or go to our web site to see YT-UGC movies and their subjective high quality information. We additionally hope that the improved YouTube-UGC dataset allows extra analysis on this area.

Acknowledgements

This work was potential by means of a collaboration spanning a number of Google groups. Key contributors embody: Balu Adsumilli, Neil Birkbeck, Joong Gon Yim from YouTube and Junjie Ke, Hossein Talebi, Peyman Milanfar from Google Research. Thanks to Ross Wolf, Jayaprasanna Jayaraman, Carena Church, and Jessie Lin for his or her contributions.