A Google web engineer has created an artificial intelligence-powered (AI) tool for browsers that can erase people from webcam feeds. If that alone isn’t cool enough, the tool can do it in real-time.
Jason Mayes developed the Disappearing-People tool with Javascript and TensorFlow, Google’s free and open-source software library that lets developers apply AI and deep learning to smaller-scale applications like this. In the case of Mayes’ tool, the neural network determines what is the static background of a video and creates a ‘clean plate’ version without anyone in it.
Then, the program recognizes people and uses that knowledge to generate a clean image of a webcam’s feed without the person. However, what makes this particularly exciting is that it does it in real-time. In other words, someone can walk into frame and move around and the webcam will continue to play live footage, but the person won’t be visible.
Unfortunately, as impressive as it is, the software isn’t completely perfect. In a sample video posted by Mayes, you can see some artifacts around him as he moves. Further, Mayes walks across a bed in the video, causing it to change shape, mess with the AI and reveal his legs. However, for all the artifacts, the manipulated webcam feed never reveals Mayes’ face.
Practically, there are a variety of uses for this software. For one, installing it could negate the need for putting sticky notes over your webcam since anyone trying to spy on you won’t actually see you. With some extra work, it could also be useful in filmmaking, especially with movies or shows that rely on puppets. The AI could erase people from the shot and reduce the amount of visual effects work needed.
Alternatively, it could be a great feature for remote workers calling into meetings via video. Again, with some tweaking, the AI could be trained to erase other people from the feed but keep you in view.
Ultimately, this could be the start of an incredibly useful tool for a great many people. It’s also available to test out if you want. Mayes shared test versions of the tool on his personal GitHub repository.
Source: Gizmodo
