Reproduce “DeepArUco++: Improved detection of square fiducial markers in challenging lighting conditions” on OpenResearch

About this paper

Rafael Berral-Soler, Rafael Muñoz-Salinas, Rafael Medina-Carnicer, Manuel J. Marín-Jiménez

DeepArUco++ presents a deep learning-based pipeline for robust detection, precise localization, and accurate decoding of square fiducial markers. The system achieves superior performance compared to traditional and existing deep learning methods, especially under challenging and non-uniform lighting conditions, by leveraging tailored synthetic data generation and a modular neural network architecture.

Key insights

• Heatmap-based corner refinement consistently outperforms direct regression for precisely locating marker corners, leading to higher rates of successfully identified markers, despite a slight increase in processing time.
• Careful data augmentation (blur, color shift, Gaussian noise) for the YOLOv8 detector and online augmentation during corner refinement training are critical for achieving robustness against diverse lighting conditions.
• While competitive on general datasets, DeepArUco++ demonstrates a substantial performance leap over classical methods and existing deep learning approaches specifically under challenging, non-uniform lighting.

Summary by alphaXiv.