This paper addresses the problem of localizing hazard-shaped calibration markers mounted on eyeglass frames, a task relevant to augmented reality, head pose estimation and optical distance calculations. The goal is to accurately detect these markers, which lie in the upper portion of the face, within a rectangular region provided by a face detector. We revisit this problem and ask: how much precision can be reliably extracted, and at what computational cost? Beyond a controlled comparison of heatmap-based segmentation models (HRNet modified for multiscale outputs, MobileNetV3) and object detection with YOLOv8 variants, we introduce two contributions that improve accuracy while retaining real-time performance: (i) a temperature-scaled soft-argmax readout for subpixel localization with calibrated spatial confidence, and (ii) a lightweight Detection → Refinement cascade that couples YOLO proposals with a 7-layer heatmap refiner operating on $64 \times 64$ crops. Experiments are conducted on both synthetic and real datasets spanning diverse capture distances. Our cascade reduces mean pixel error by 17 % over YOLOv8-n while running at $\mathbf{7 8 ~ m s}$ per frame (includes additional latency such as image loading, resizing, and patch extraction, inference time alone is 75 ms per frame), and outperforms HRNet by 4.7% AUC@2px at a fraction of its latency. Evaluation includes distance-based error metrics, uncertainty calibration (ECE), and both quantitative and qualitative analyses. Results show a strong correlation between confidence and localization error, enabling practical deployment.