Why measuring map quality matters more than you think

Autonomous driving systems predict road maps as sequences of points, but existing evaluation methods (Chamfer distance + mAP) ignore point order and hide geometric flaws. This work introduces SOSPA, an order-aware metric for individual map elements, and PLD, a soft detection framework that replaces hard thresholding with principled scoring. Tests on nuScenes show PLD correctly ranks leading methods and exposes detection capability—not geometry quality—as the real bottleneck. Code will be released.