Monocular Object Dimension Estimation via YOLOv8 and Embedded A4 Reference Calibration Without Depth Sensors

Measuring real-world object dimensions from a single monocular image is harder than it looks. Without depth data or explicit camera calibration, most systems either rely on dedicated hardware or make assumptions that break down in practice. We present a fully automated deep learning framework that sidesteps these constraints by using an ordinary ISO A4 sheet as an in-scene reference. The system detects both the A4 sheet and the target object in a single pass using YOLOv8, derives per-axis pixel-to-centimeter scale factors from the sheet’s known dimensions (21 cm × 29.7 cm), and outputs width, height, and diagonal measurements (Eq. 1) without any additional sensors or manual steps. The implementation runs in Python 3.10 with the Ultralytics API, OpenCV for image processing, and a Tkinter GUI. We built and annotated a custom dataset of 2,240 images across nine object classes, then augmented it to 6,720 samples for training. The model achieves mAP@0.5 of 97.46% on the A4 reference class and 96.99% across all classes, with an average measurement error of 0.18 cm and GPU inference under 20 ms — outperforming all seven compared state-of-the-art methods on both detection accuracy and measurement precision.