DeepIRTag: A Framework for Invisible Fiducial Marker Design and Detection

This paper introduces DeepIRTag, a novel framework for the design and detection of invisible fiducial markers to resolve the visual incongruity of traditional high-contrast markers in augmented reality (AR) and robotic navigation. Leveraging infrared-reflective materials and dynamic grid-based Hamming encoding, the proposed markers remain imperceptible under visible light while enabling robust detection in infrared imaging. The framework integrates a convolutional neural network (CNN) architecture enhanced with two key modules: the Universal Inverted Bottleneck (UIB) for efficient global context capture and computational optimization, and the Context-Aware Receptive Field Fusion (CARUF) with a Contextual Anchor Attention (CAA) mechanism to improve multi-scale feature fusion. Evaluated on synthetic datasets simulating real-world scenarios (e.g., low illumination, occlusion), DeepIRTag achieves state-of-the-art performance, with 97.8% ROI-mAP50 and 96.9% KPT-mAP50, surpassing conventional markers like AprilTag and RuneTag. In challenging environments (distance >50 cm or viewing angles >60°), it maintains robust detection (60% precision, 48.97% recall), outperforming existing methods. Ablation studies confirm the synergistic benefits of UIB and CARUF, demonstrating a 5.3% improvement in keypoint detection over baseline models. This work advances fiducial marker technology by harmonizing environmental aesthetics with machine perception, offering scalable solutions for applications requiring unobtrusive yet reliable spatial tracking.