We all watched a picture at a given moment of something like the ocean, the clouds or the waterfall, and for the shortest time, it almost seems that the photo moves. Typically, this perceived movement is just a thing of our mind. The researchers at the University of Washington have developed a new method of deep learning that can animate some parts of a photo, turning it into a video.
UW-deep learning methods can turn on flowing materials such as waterfalls, smoke, or clouds. The technique developed at the University only requires one photo of a waterfall to create animation. The researchers used a short video that loops smoothly, giving the impression of a sustainable water movement.
It’s almost like it can change any photos into something similar to Apple’s direct photos that capture the second or more movement before the still image. The researchers about the project said that what was unique about their method was that it did not require user input or additional information. The process only requires images and produces high resolution videos, smooth looping which usually looks like a real video.
Developing technology is a challenge. According to the researchers, it effectively requires them to predict the future. This system contains two parts, with the first part that predicts how things move when the photo is taken. The information was taken and used to create animations.
Motion estimates require researchers to train neural networks with thousands of waterfall videos, rivers, oceans, and other materials with fluid movements. The training process requests a neural network to guess the video movement when only given the first frame. Neural networks compare predictions with actual videos and learn to identify instructions, such as ripple in flow, to help predict what happens next. The researchers created something they called “symmetrical sparks” that predict the future and past for an image, combining it into an animation.
