Video is ubiquitous online. Estimates say 75% of all traffic is video content, with this number expected to triple by 2021. MIT Technology Review reports how deep learning is innovating video compression.
Advances in compression techniques have been stagnant for years. “The fundamentals of existing video compression algorithms have not changed considerably over the last 20 years,” Oren Rippel and his team at WaveOne, a deep-learning startup, told MIT Technology Review.
Their new deep learning algorithm offers a superior way to compress video. “To our knowledge, this is the first machine learning–based method to do so,” they said to MIT Technology Review.
When software compresses video, it takes out redundant data from a code and inserts a shorter description that still enables a later reproduction. As MIT Technology Review explains, this is usually a two step process.
Motion compression is the first step and it involves scanning the video for moving objects and predicting future object movement to encode only the object shape and direction of travel rather the associated pixels for the object in motion per frame.
Completing the process is residual compression. This step eliminates repeated data between frames. For example, an algorithm could recognize that pixels of a certain area of a frame do not change.
Rippel and his team are improving the process by applying machine learning to both steps. With motion compression, their machine-learning techniques have uncovered new redundancies that standard codecs would miss.
Their approach also sidesteps the usual headache of having to determine how bandwidth should be allocated for motion and residual compression. Traditional compression algorithms conduct these processes on a separate basis.
WaveOne’s solution compresses both signals simultaneously and determines how bandwidth should be allocated based on frame complexity.
Their resulting compression algorithm surpasses traditional video compression codecs, significantly cutting down size and download times for online vide, both HD and standard definition.
While the team still has some efficiency gains to achieve since their decoder is still on slower side, they see their current product as a proof-of-principle. “The current speed is not sufficient for real-time deployment, but is to be substantially improved in future work,” they say.