How a printed pattern can fool a person detector
A camera does not recognize a person the way a human does. YOLOv2, the object-detection algorithm used in this demonstration, processes an image and predicts where objects are and what they might be. It assigns confidence scores to those predictions; if the score for “person” falls below the system’s threshold, the detector may omit the person from its output.
The colorful patch is designed to disrupt that process. Its pixels are chosen through repeated tests against the model: the designers adjust the pattern until the detector becomes less likely to identify a person when the patch appears in the image. To a viewer, the result can look like random psychedelic art. To the algorithm, it creates image features that interfere with the features it normally uses to detect a body.
The patch attacks the model, not the camera
The pattern does not switch off the camera or erase the person from the recording. It targets the recognition stage—the software that labels objects in the footage. In the demonstration, someone holds the printed patch in front of their body, and the detector can fail to produce a person detection. That can make the person seem “invisible” to that particular automated system while leaving them plainly visible to a human watching the video.
The distinction matters because the result depends on the model and the conditions. A patch optimized for YOLOv2 is not a universal cloak: another detector, a different viewing angle, lighting, distance, or camera setup may produce a different outcome. Nor does defeating one object detector prevent other forms of monitoring or analysis.
The revealing point is how a small, physical change to an image can exploit a weakness in a system that appears to understand what it sees. A camera may record the whole scene, yet the software tasked with finding people can miss the person standing directly in front of it.
