Techniques for reliably fooling AI machine-vision classifiers

The Open AI researchers were intrigued by a claim that self-driving cars would be intrinsically hard to fool (tricking them into sudden braking maneuvers, say), because "they capture images from multiple scales, angles, perspectives, and the like." Read the rest

Machine vision framework wants you to put down your weapon

This is doing the viral rounds described as a Google technology, but it's actually Apple's VisionCore in action. It runs offline on the local device, requiring no number-crunching help from the cloud. Here's a breakdown of how it identifies things through code.

You will need the beta version of xCode and a device running the iOS 11 beta (make sure you only install the beta software on a test device!).

I liked watching it contemplate whether a metal ruler was a meat cleaver or a "chopper." Whispers of the ACLU lawsuits of tomorrow: I think you'd better do what he says, Mr. Kinney. Read the rest

More Google Earth anomalies

Artist Clement Valla collects the most remarkable machine-vision nightmares and curiosities from Google Earth, a world whose parallels to our own become uncannier with each sweep of the satellites and Googlecars. [Previously. via] Read the rest

Glitché

Glitché is the evil twin of all those old-film, toy-lens, Instagram-style apps. Pick a photo, then glitch it all to Hell with broken NTSC emulation, weird 3D pixelation and heightmap extrusion effects, and delicious MPEG-style compression errors. For a $1 upgrade, the free app lets you save animated GIFs, too. [via Joel Johnson, below] Read the rest