Researchers at NEC Laboratories America and Duke University have developed an algorithm that creates videos from text snippets.
The first stage of the process involves a "generator" neural network using text to produce a blurry image of the background with an unfocused blob where the main action occurs. The second stage derives a video from both this "gist" and the text, producing a short video.
A second network functions as a "discriminator" during training by watching the generated video alongside a real video of the action described in the text, and is taught to pick the real one. This network's critical ability improves over time, with its feedback setting a higher standard for the generator network.
The team trained the algorithm on 10 types of scenes, and a simple classification algorithm correctly guessed the intended action among six choices about 50% of the time.
View Full Article
Abstracts Copyright © 2018 Information Inc., Bethesda, Maryland, USA
No entries found