Hive AI Automates Image Recognition for Media Companies

Jun 16, 2017 | 5124 Views

Machine learning and image recognition companies often require loads of data in order to fine tune their software to work well.

Castle Global decided to solve this problem by generating its own data. The startup built a number of consumer apps, including chat app Chatous, Q&A app Kiwi, local classifieds app Plaza, and Melon, an app for meeting and chatting with people.

These apps, with about 50 million users combined, provide the company with reams of data.

Castle Global then built an internal product to identify objects within images in its apps--for things like offensive content.

That technology became Hive AI, which the company has launched and sells to a variety of media companies. It hopes to expand to other verticals. Customers include GIF company Tenor,, a social app in Europe, and unnamed large media companies.

Castle Global, with 90 people, has raised $15 million, including a $13 million Series A led by 8VC with participation from Founders Fund and General Catalyst, as well as a $2 million seed led by General Catalyst.

Media companies can use Hive AI's API to identify celebrities, actors, musicians and politicians, and identify brands or other objects within images or videos.

Media companies can use Hive AI to identify when viewership increases or decreases during a television show, news show or television or video ad--based on which people are on screen. It's also used for user-generated content.

Automatically extracting this type of data from large stores of photos or videos is not easy, but Hive can do this quickly, said Kevin Guo, co-founder and chief executive at Castle Global.

"We go frame by frame and find this scene with this person who was on for 3 minutes and see what happened to viewership," Mr Guo said. "Typically it's hard to tag that manually. We can find every logo present. We dynamically do all this to find trends and patterns."

The company initially built the product by building its own form of Mechanical Turk with workers to identify objects, such as a dog or a cat. Then it uses that data with machine learning for Hive to identify objects on its own.

This technology for labeling data is just as critical as the large amount of data it has amassed to identify images, Mr. Guo said.

"It surprised us, to be candid, how powerful the technology has been and the accuracy and breadth of its use cases," said Alex Kolicich, a member of the investment team at 8VC. Continue Reading>>

Source: WSJ