For a human being it is easy to predict what will happen immediately, observing for example the movement of an object. For a car it is impossible, but scientists of MIT developed an algorithm of deep learning which allows you to create a “ future video ” from images of a static scene. A similar artificial intelligence could be used for security cameras in the car or autonomous driving.

The system designed by the Computer Science and Artificial Intelligence Laboratory (CSAIL) required a long training with 2 million videos. Instead of creating frame-by-frame scenes, the algorithm generates 32 frames in a second separating the background from the object in the foreground. The model is thus able to distinguish which items are firm and what the move. The researchers used a technique called “adversarial learning” which provides for the training of two neural networks which generates a video and another that checks whether the video is real or simulated.

You get the final result, that is, the future video when the first neural network is able to fool the second, which considers real video simulated. The system can then predict the waves crashing on the beach or people walking on the grass. The videos generated were also considered real by people who participated in the test.

The system is not yet perfect. There are indeed some limits, such as the maximum video length (1.5 seconds), but the MIT team hopes to increase the length in the future. A similar artificial intelligence can be exploited in several ways. You can, for example, add animations to still images achieving an effect like that of the Prophet in the Harry Potter films. A self-driving car may instead predict the movement of other cars, cyclists and pedestrians.

SHARE

LEAVE A REPLY