DeepTag: A General Framework for Fiducial Marker Design and Detection

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Traditional fiducial marker systems rely on hand-crafted detection algorithms limited by low-level image processing.
• Existing methods struggle with marker appearance variations and require complex coding systems to compensate for detection algorithm shortcomings.

Purpose of the Study:

• To introduce DeepTag, a general deep learning framework for fiducial marker design and detection.
• To enhance flexibility and robustness in fiducial marker applications.
• To enable the design of novel marker families with customizable patterns.

Main Methods:

• Developed a deep learning-based framework, DeepTag, for fiducial marker detection and design.
• Implemented on-the-fly synthetic training data generation, eliminating the need for manual annotations.
• Collected a new large and challenging dataset with diverse marker view distances and angles for validation.

Main Results:

• DeepTag demonstrates robust detection across various existing marker families.
• The framework enables the creation of new marker families with customized local patterns.
• Experiments show DeepTag significantly outperforms existing methods in detection robustness and pose accuracy.

Conclusions:

• DeepTag provides a flexible and robust solution for fiducial marker detection and design.
• The deep learning approach overcomes limitations of traditional image processing techniques.
• DeepTag facilitates adaptation to both existing and newly designed marker families, advancing computer vision applications.