Like Silicon Valley’s plot on fictional startup Pied Piper, Compression AI is a scrappy team of developers working on media-compression technology in a tech incubator.
Except instead of being characters in a Hollywood-scripted startup, founders Francis Doumet and Migel Tissera met at a Vancouver coworking space, hired two employees and pulled late nighters to release their first beta software, dubbed PixelDrive.
Founded in 2018, Compression AI aims to enable faster transmission of media files over the web, even on low-quality internet networks.
It reduces image file sizes up to 80 percent using the company’s neural compression technology, the product of the team’s custom work on convolutional neural networks trained on NVIDIA GPUs.
Doumet and Tissera initially launched PixelDrive as a consumer product. But they soon figured out that the underlying technology is much more valuable to developers because image compression enables faster web page load times and increases search engine rankings. They have since made the technology available as an API for developers.
Doumet and Tissera’s ultimate goal is to bring their technology to video compression. That’s because Tissera — a fan of watching UFC mixed martial arts fights but frustrated with choppy broadcasts — sees a need for improvement in video compression, especially where the internet quality is suboptimal.
Compression AI is a member of NVIDIA Inception, a virtual accelerator program that helps startups get to market faster.
Compressed Launch Date
The neural networks that run the developer API and PixelDrive were trained on the entire ImageNet set of images and many more that were collected from the web, totaling more than 10 million images, Doumet said.
The Compression AI team designed the neural networks, which focused on the part of CNNs known as auto-encoders, he said. The development allowed Compression AI to come up with the optimal image compression for each individual image down to the pixel, according to the company.
The deployed service is powered by NVIDIA P4 GPUs performing inference in the cloud. “We’re best in class in terms of image compression,” said Doumet.
Neural network training was also fast on desktop PCs running NVIDIA GPUs.
Online Business Applications
Improved image compression has potentially big implications for businesses. The startup has multiple pilot tests with companies exploring the benefits.
One is with a major online real estate site. Sites like these rank higher in Google searches if they load faster from better compression of images, said Doumet.
Another is video game apps because lighter file sizes from image compression get lower bounce rates at the time of download.
And online retailers are exploring pilots to get better sales results from fast load times of pages, according to Doumet.
Coming Attraction: Video
Compression AI is focused on launching compression for video next.
Doumet and Tissera say that even with advances in 4G and the promise of 5G, mobile internet remains bandwidth constrained. For instance, a four-minute video shot in 4K on a mobile device takes roughly 13 minutes to transmit over the average U.S. internet connection.
“Advances in AI create an opportunity to develop more intelligent kinds of codecs that can adapt and optimize for any image to offer a reduced footprint in file size,” said Doumet.
Image license and credit: Creative Commons; photo by @noisytoy.net