Industrial organisations involved

KEBULA is a Data Lab born as a spin-off company of the University of Salerno, qualified as innovative start-up. The mission of KEBULA is unlocking the full potential of data to ignite tomorrow’s business.
KEBULA helps its clients to enable their data journey and enhancing their data engineering, analytics and AI capabilities.

KEBULA is a founding member of MPAI (Moving Pictures and Data Coding with AI), which sets new standards for encoding multimedia data using Artificial Intelligence (AI).

Technical/scientific Challenge

Video streaming is now ubiquitous and fundamental for numerous business applications, including web streaming, remote sensing and satellite monitoring.

Current efforts are being made into designing new video codecs that can reduce the bitrate required for storage and transmission of high-quality videos.

The use of AI is a promising direction in this regard. In particular, super-resolution models can enable transmission and storage of low-resolution videos, which can then be upscaled to the desired resolution without a significant loss of quality intrinsic to traditional upscaling methods.

The Solution

Different models were evaluated using the standard BD-PSNR and BD-RATE.

All models were compared with the HEVC codec, tuned for streaming settings. The largest models recorded the best super-resolution quality: IconVSR obtained 1.127 BD-PSNR and -28.511% BD-RATE when computed on the 3 RGB channels. Thus, on average, upscaling 1080p videos to 4K resolution using IconVSR would produce a 1.127 higher PSNR at the same bitrate. EDVR also obtained similarly good results, with 1.101 BD-PSNR and -27.759% BD-RATE.

However, these models could only run at 0.79 and 0.73 FPS on Marconi100 respectively, making them unsuitable for real-time applications. On the contrary, smaller models proved promising for real-time applications. For example, the base EVSR model can run at 82.03 FPS on Marconi100.

Business Impact

The project has had several positive effects for KEBULA.

First, KEBULA gained experience on performing multi-node, multi-GPU training.

Second, KEBULA acquired knowledge and experience on the usage of video super-resolution models. Super-resolution has been gaining traction in recent years since it can be applied to a variety of applications beyond video compression. As more powerful GPUs will be available and existing models improve, more possibilities for real-world applications of super-resolution will open.

Since KEBULA aims to cover a wide range of data analysis and transformation pipelines, experience on this kind of models will be more and more relevant to open new business opportunities.

Third, since the PoC was born in the framework of our collaboration with MPAI regarding the design of new, better, video codecs enhanced by AI algorithms, the knowledge acquired by testing and evaluating all the described models will concretely help KEBULA to determine possible paths to reach this goal.

The benefits

  • The models developed paves KEBULA’s current and future work with MPAI for video compression algorithms.
  • The super-resolution algorithms implemented will be used by KEBULA in the aerospace domain, where applications have stricter hardware and power constraints.

Images Courtesy: Kebula