YouTube channel Deepython published a video where a neural network based system with high accuracy recognizes people and objects on the streets of New York.
Users in the comments cannot decide whether to admire the development of technology or already start to fear mass surveillance.
Behind the Deepython channel and the site with the same name is the 22-year-old Clayton Blythe, a programmer and a computer training specialist. In the description of the video, he says that he used the taken online video walk through New York and processed the video data using the Faster R-CNN architecture-based neural network. The technology allows you to determine with a high degree of precision where a particular object is on the image, and with the help of the library Tensorflow recognizes individuals in the crowd, distinguishes trucks from cars, and bags from backpacks.
The type of each object appearing in the frame of the object is recognized with a certain degree of accuracy, up to 99 percent.
In the development of the technologies used, the author did not participate in the video, and used the pre-recorded footage of the walk through Times Square to create the video. But there are already systems that can determine the type of objects in real time — true, to the detriment of accuracy.
Sometimes there are minor disruptions. For example, at the 16th second the video program for a split second takes the woman walking as if she’s the horse.
A bottle of milk in the hand of a passer-by for some reason confuses with a frisbee dish.
And the mannequins in the showcase have so far been given the status of people with a 60-70 percent probability.
Neural networks have learned not only to recognize objects and people, but also to create realistic faces based on celebrity photos.
Demonstrating video technology appeared on the channel of deepython.com on November 22 and gathered more than 80,000 views.