Creative agency Redpepper made a AI-powered robot that can pinpoint Waldo in 4.45 seconds (“better than most five year olds,” according to Redpepper). The robot is complete with a rubber hand that points to Waldo on the page.
The agency used Google’s AutoML Vision service to train AI on photos of Waldo. The drag-and-drop tool allows users to train AI tools without previous coding knowledge, and has been used to categorize anything from ramen based on the shops they came from to the types of attire carried in an online retailer.
Matt Reed, a Creative Technologist at Redpepper who led the project, got 62 Waldo heads and 45 full-body Waldos from Google image search, then fed the data Google’s AutoML Vision. “I thought that wouldn’t be enough data to build a strong model but it gives surprisingly good predictions on Waldos that weren’t in the original training set,” he told The Verge.
The AI is paired with a metal robotic arm, a Raspberry-pi controlled uArm Swift Pro with a Vision Camera Kit to perform facial recognition. The camera snaps photos of pages in Find Waldo books, using OpenCV to detect possible Waldo faces in the photo. Then, the images get sent to Google’s AutoML. When the AI determines a match with 95 percent confidence or higher, the rubber hand points out a Waldo.
This robot completely defeats the purpose of Where’s Waldo, of course. The AI can’t appreciate the infuriating/gratifying purpose of Waldo searching, or admire particularly funny details in an illustration. Reed had some other ideas for the same system: “Maybe a fun use would be seeing what cartoon character the AI thinks you look closest to? Maybe could detect comic book forgeries?” Maybe it can even be adopted to ruin another beloved children’s puzzle?
(source: The Verge)