Efficient HD Map Generation for Autonomous Driving With Spline

Abstract

This study proposes a novel framework for generating vectorized high-definition (HD) maps with maximized representational efficiency from multiview images. While existing end-to-end methods have shown promise, they often predict a fixed, dense set of points, leading to data redundancy and difficulties in representing smooth, continuous curves. To address this limitation, we adapt herein the Mask2Map framework and introduce a new methodology within its vectorization phase (multistage manufacturing process network). The core idea is to treat the predicted vector points as control points (not as a final shape) for a continuous Catmull– Rom spline. We chose the Catmull–Rom spline owing to its unique property of passing through all the control points, making the relationship between prediction and the final curve highly intuitive. This approach facilitates the learning of a highly compact and efficient representation. The primary objective is to maximize this efficiency rather than peak performance. Experiments on the nuScenes dataset demonstrate the success of our method, which maintains a comparable mean average precision (mAP) of 69.83% while utilizing only half the control points (10 points) required by the point-based baseline (70.83% mAP with 20 points). This significant reduction in data density with only a marginal performance trade-off validates the superior efficiency and practicality of the proposed spline-based representation for HD map generation.