InterDigital develops CompressAI platform
September 17, 2020
InterDigital, a mobile and video technology research and development company, has developed, CompressAI, a software platform to simplify and accelerate AI-based video compression research. CompressAI allows researchers to quickly design, train, test, and evaluate AI-based codecs. The open source platform, recently released to the video compression research and standards communities, is already being used to support the development of next generation image and video codecs.
Research teams around the globe are examining how AI-based approaches, in particular deep learning and neural networks, can be used in the next generation of image and video codecs. However, the process to complete AI-based video compression research often includes several tedious and time consuming steps. Researchers who invent a new neural network component or architecture must typically augment this component with other neural network layers to make a fully functioning image or video compression network, then find image and video sequences with which to train their network, and compare the results with existing compression standards. Moreover, this process must be repeated every time a team develops a new AI-based codec to evaluate.
Leveraging more than two decades of experience in video compression and the independent research of InterDigital’s AI Lab, CompressAI automates a large proportion of this research process to accelerate the development and evaluation of candidate neural network architectures in deep video compression research. The CompressAI platform features the following software modules:
- Data loader: Loads data used to train neural networks. The data loader includes pointers to several publicly available data sets that can be used for training purposes.
- Data formatting: Translates and formats training data (which is made of images and video segments stored in a wide variety of file formats and binary representations) into a simple representation which can easily be understood by neural networks.
- Reference layers: Includes high-performance implementation of several types of specialized neural network layers used in video coding, such as layers that implement entropy coding or generalized divisive normalization.
- Pretrained models: Includes implementations developed by InterDigital, as well as downloadable pretrained weights of several (already published) state-of-the-art neural networks for deep image and video compression.
- Benchmarking: Compares the performance of both traditional and deep learning-based codecs using well known objective metrics such as PSNR (Peak Signal to Noise Ratio) or SSIM (Structural Similarity Index Measure).
“The inexorable rise of rich media and high-quality video streaming is putting significant strain on our networks, which means continually improving codecs’ compression capabilities is a must. Many research teams across the world are now looking at the role of AI-based codecs and are developing and evaluating new potential standards to meet the ever-growing demands for video content. But developing these AI-based codecs involves several complex, time-consuming and tedious steps,” said Jean Bolot, Vice President of the AI Lab at InterDigital. “InterDigital has developed the CompressAI platform to simplify and speed up the process of deep video compression research, allowing teams to test and evaluate AI-based codecs quickly and accurately. By making our platform publicly available, we hope to be able to support the incredibly important work research teams are doing to bring the next generation of video compression standards.”