Learnable Visual Markers

TL;DR

This paper introduces a novel method for designing visual markers using deep generative networks, enabling adaptive, stylized, and robust markers for augmented reality and robotics that retain information despite distortions.

Contribution

The authors develop a joint training framework for synthesizing and recognizing visual markers with deep networks, incorporating stylization and distortion robustness, advancing marker design capabilities.

Findings

Markers retain long enough bit strings for practical use

Markers can be stylized to match texture prototypes

The approach offers insights into pattern recognition by ConvNets

Abstract

We propose a new approach to designing visual markers (analogous to QR-codes, markers for augmented reality, and robotic fiducial tags) based on the advances in deep generative networks. In our approach, the markers are obtained as color images synthesized by a deep network from input bit strings, whereas another deep network is trained to recover the bit strings back from the photos of these markers. The two networks are trained simultaneously in a joint backpropagation process that takes characteristic photometric and geometric distortions associated with marker fabrication and marker scanning into account. Additionally, a stylization loss based on statistics of activations in a pretrained classification network can be inserted into the learning in order to shift the marker appearance towards some texture prototype. In the experiments, we demonstrate that the markers obtained using our approach are capable of retaining bit strings that are long enough to be practical. The ability to automatically adapt markers according to the usage scenario and the desired capacity as well as the ability to combine information encoding with artistic stylization are the unique properties of our approach. As a byproduct, our approach provides an insight on the structure of patterns that are most suitable for recognition by ConvNets and on their ability to distinguish composite patterns.