Researchers with GPU company, NVIDIA have created an AI program based on deep learning that’s able to create smooth slow-motion videos using automated frame interpolation, and even slow down existing videos even further. The program uses deep learning with a convolutional neural network in order to estimate what’s taking place between frames through object identification and comparison of sets of two frames for each step. It then interjects what it thinks should go between two frames, and eventually creates a smooth slow-motion video from a regular piece of footage. NVIDIA’s researchers have not yet revealed just how the whole thing works in detail, but they plan to show it off at this week’s Computer Vision and Pattern Recognition conference.
NVIDIA’s new solution is reportedly able to outperform a number of comparable solutions that are currently leading the market, and is scalable to almost any application thanks to its use of a convolutional neural network. It was trained on over 11,000 different videos. To put the program to the test, NVIDIA took a variety of footage from around the web and slowed it down, then took some videos from YouTubers The Slow-Mo Guys and slowed them down even further. All of that can be seen in the video attached at the end of the article, including such mesmerizing spectacles as jelly hitting a tennis racket and a man jumping on and bursting a giant balloon.
Slow motion video is impractical to shoot every moment in, due to its processor intensity on even modern devices, and the large amount of storage it takes up. This program has the potential to change all that by offering users the chance to slow down videos after they’ve already been taken. NVIDIA obviously has yet to announce any plans as to how it may offer this technology to the public, so don’t expect to be able to slow down old family videos smoothly using NVIDIA’s AI and processing grunt just yet. The program relies on large clusters of powerful computers to run, which means that it will likely have to be offered as a cloud service in some form.