Academia|Photos want to be still but AI doesn't, MIT black tech turns images into small videos in seconds
Have you ever thought about it? When you are shown any photo, you may see not just a still image, but a moving "mini-video". Today, with the help of machine learning, it is possible to predict the next sequence of actions based on a still photo, with a fairly high accuracy rate.
Whether it's a beautiful woman riding a bike, a dog catching a Frisbee, or someone falling suddenly, etc., imagining these continuous actions is one of our most basic skills, and we don't need to consider the vast amount of information used to predict them, such as gravity, inertia, and the instinctive response to a fall. Then, to get the computer to learn this ability to anticipate is certainly a key challenge in machine vision.
Researchers from the Massachusetts Institute of Technology are working to solve this problem, and they have shown a series of very impressive results. By using specially trained neural network , converting the image into a video and having the computer predict what will happen next. However, their model still has many limitations; the videos are usually only a few seconds long, the files are small, and the images are often messy. But it's still an impressive creation in machine imagination, and computers have taken another step forward in understanding the world as humans do.
Train this. neural network The use of more than 200 10,000 from Flickr Downloaded video clips。 All scenes are divided into four types： golf course、 beaches、 Train stations and hospitals。 That's a solid set of continuous shots.， Eliminates camera shake。 With these data， Team neural network Not only can you generate short videos of scenes like these， Can also produce a continuous picture from a still image。 This is essentially a prediction of the action that will happen next， However, the current effect is still very limited， One can only speculate on the pixel change， Instead of an understanding based on the whole scenario。
Here are the results.