2019年11月15日
By 1月时特约编辑
专题文章

3 Takeaways From Moscow State University's 2019 Codec Comparison

在我看来, one of the most useful and credible codec analyses comes from Moscow State University, 它刚刚发布了 2019视频编解码器比较. This essential resource includes multiple HEVC codecs, plus VP9 and AV1. 一如既往地, you’ll have to pay to get all the data ($760 through 11/30/2109, 950美元之后), but the free version provides useful data points. There are two versions of the report; the Main report which tested 12 codecs with 100 videos using objective metrics, 主观报告, which tested 5 clips and 11 encoders subjectively rated by 732 viewers.

The Main report tests three use cases—fast encoding (60 fps), 通用编码(25fps), and ripping encoding (1 fps)—and if the codec doesn’t meet that minimum performance threshold it’s not considered for that test. The Subjective report only considered codecs that could produce 1 fps. 2020年1月, MSU plans to release a report considering high-quality codecs that can produce a frame every 2 minutes.

Here are three takeaways from the 2019 MSU report.

x.265可能不是你最好的选择

x265 is free with FFmpeg so it's the most accessible HEVC codec, but it has come up short in multiple MSU reports and this one is no exception. 通过解释, all three charts below use x264 performance as the benchmark, 所以它总是100%的. Lower numbers indicate the ability to produce the same quality as x264 at the percentage shown, so the highest-performing HEVC codec in 图1, HW265, can deliver the same quality as x264 at 57% the data rate. x265 requires 76%, which is 33% higher than HW265. So, 如果你正在考虑用HEVC编码, it’s worth exploring the other higher-performing HEVC codecs shown in the report.

图1. 客观表现以SSIM衡量.

AV1在实际测试中看起来更好

Like HEVC, AV1 is a standard so there will be multiple codec versions. WZAurora AV1编码器(来自 Visionular) tested by MSU not only met the 1 fps speed requirements but also produced dazzling quality (as shown in 图2), outperforming the best HEVC encoder by about 28%.

图2. Subjective performance of multiple codecs as measured by Subjectify.us.

但是，有一个重要的警告

The sharp-eyed reader will have noted WZ Aurora in the middle of the pack in 图1 but at the front in 图2. 有什么区别?? 请看图3, which shows performance as measured by subjective testing (in red) and with SSIM (in blue). In all relevant cases save the WZAurora AV1 Encoder the results are fairly consistent. With WZAurora, the differential was nearly 50%.

图3. 客观的对. 测试编解码器的主观评分.

我问密歇根州立大学的小组组长Dr. Dmitriy Vatolin about this scoring discrepancy. He responded, “Currently, we can’t give the exact answer, but we can guess some possible reasons. 例如, it can be contrast or brightness increase of encoded stream, 对于观察者来说，哪一个看起来更好. We did such a study when we analyzed the applicability of the VMAF metric to objective comparisons (see our paper 'Hacking VMAF With Video Color and Contrast Distortion.'"

I’m hoping to test the WZAurora codec in early 2020, and this is an issue I can look for. 与此同时, the MSU report provides lots of valuable data about the real-world performance of multiple HEVC codecs as well as AV1 and VP9.