The Google algorithm has learned to recognize objects on video





"Find me a movie with the same funny dog." Launched in beta mode, the Google Cloud Video Intelligence service will suggest: to start a search for the dog's breed, its size, the length of the coat or the stupid expression of the muzzle? For a new system, this is not a set of pixels in the picture, but a complex and important object. Like everything else in this video.


The new service is built on the basis of the Tensorflow project using the principles of machine learning. The goal is to learn how to recognize any video content based on its content so that it can later effectively search for relevant queries. Be it small, specialized fragments or large films in their entirety.


What was originally a whole video image, after processing is divided into an array of individual objects with nominal and verbal labels. They are given a weight or a rank, in percentage terms, which is formed on the basis of comparisons with similar requests. The information is taken from the usual search queries, and the result of the check is used to increase the relevance of new issues.

The more accurately placed tags, the higher the chance to find the right video, but Google is tactfully silent about the mechanisms for controlling this process. On the contrary, according to the leading specialist of the machine training corporation, Fei Fei Li, this API is designed for large businesses, media holdings and service providers. Those who need an effective way to manage content. In its purely commercial purposes, of course.


In the current format, innovation is in no way suitable for implementation in custom products, everyday applications. Too cumbersome and "stupid". However, the trouble is the beginning and the technology of search for video content itself is likely to become a key tool for working on the Internet in the near future.

Comments