It is important to stress that apart from assigning it an orange label (instead blue) in the diagrams below, our own algorithm got no other special privileges during testing.Įxample: Single frame of BCC UpRez x3 upscaling extracted from the “Dinner Table” video at 00:00:03 Test MethodologyĪll upscalers were tested on three downsampled Shutterstock 1080p HD videos via FFmpeg using the VMAF metric. It is able to perform accurate upscaling up to x4 and produces a predictable, robust output which is a distinct advantage over e.g. “ PIXOP Super Resolution” (internally known as “PABSR1” and part of PIXOP Platform available on is our machine learning based algorithm trained on tens of millions of low-res/high-res pairs of image patches. Significant consideration and experimentation was put into aligning the settings as much as possible. Some products offer more flexibility than others. Where applicable screenshots of the settings used are included below. It’s outside the scope of this post to dive deeply into the technical details of each. Each algorithm obviously has its own strengths and weaknesses in areas such as visual quality, speed, ease of use, and flexibility from a UI standpoint. Both of these were obviously included.ĭue to the proprietary nature of many of the algorithms involved it’s impossible to make any predictions about their complexity and sophistication. and Adobe After Effect’s Detail-Preserving Upscale. For instance, we have heard the term “industry standard in the broadcast industry” used for both Grass Valley’s Alchemist Ph.C. Initially we performed some research and identified the group of candidates to pit against our algorithm. We had to limit the scope of our testing and be smart in terms of selecting video footage and competing algorithms. The advantages of this approach are straight-forward objective quantification of performance and reproducibility as opposed to performing visual subjective testing on humans. how well challenging features such as edges, textures and curves are preserved and reconstructed. The ability to hallucinate pixels of the original video accurately is thus determined, i.e. Specifically we measure an algorithm’s ability to reconstruct video in terms of how appealing it is to a human observer via Netflix’s VMAF metric. As no relevant published benchmarks were found, we decided to conduct our own testing.
0 Comments
Leave a Reply. |