Stealthy location tracking with ninja codes

Author

Imoto, Yusuke and Kato, Shunya and Takeuchi, Yuichiro

Abstract

In this paper we describe Ninja Codes, neurally-generated fiducial markers that blend naturally into real-world environments. Taking inspiration from recent advances in deep steganography, we train a neural network that transforms arbitrary images into functional fiducial markers (Ninja Codes) with minimal visual changes. Ninja Codes can be pasted onto various surfaces, to provide location tracking capability for applications such as augmented reality, robotics, etc. Ninja Codes can be printed using common color printers on regular paper, and can be detected by any hardware equipped with a standard RGB camera and capable of running inference.