RoadPainter: Points Are Ideal Navigators for Topology transformER

TL;DR

RoadPainter tackles accurate lane topology reasoning by integrating point-based centerline regression with mask-guided refinement in an end-to-end transformer framework. It introduces a real–virtual separation in centerline queries, a points-guided mask generation module, and a mask-based points fusion step, all optionally augmented by an SD map to extend reasoning beyond the visual field. The method achieves state-of-the-art results on OpenLane-V2, demonstrating strong gains in centerline detection and topology metrics, with ablations confirming the impact of each component. This approach provides a practical path toward robust online lane topology construction for autonomous driving, reducing reliance on post-processing and improving performance in challenging scenarios such as high-curvature lanes and intersections.

Abstract

Topology reasoning aims to provide a precise understanding of road scenes, enabling autonomous systems to identify safe and efficient routes. In this paper, we present RoadPainter, an innovative approach for detecting and reasoning the topology of lane centerlines using multi-view images. The core concept behind RoadPainter is to extract a set of points from each centerline mask to improve the accuracy of centerline prediction. We start by implementing a transformer decoder that integrates a hybrid attention mechanism and a real-virtual separation strategy to predict coarse lane centerlines and establish topological associations. Then, we generate centerline instance masks guided by the centerline points from the transformer decoder. Moreover, we derive an additional set of points from each mask and combine them with previously detected centerline points for further refinement. Additionally, we introduce an optional module that incorporates a Standard Definition (SD) map to further optimize centerline detection and enhance topological reasoning performance. Experimental evaluations on the OpenLane-V2 dataset demonstrate the state-of-the-art performance of RoadPainter.

Figures (5)

• Figure 1: The illustration of our method. (a) Our model takes multi-view images as input. (b) From the constructed BEV features, two types of lane centerlines are derived: points and mask. (c) The initially detected points provide a rough localization, while the mask points enhance geometric details by utilizing the corresponding fine-grained heatmap.
• Figure 2: The architecture of RoadPainter consists of three main parts. In the BEV perception stage, we convert sensor inputs and an optional SD Map into enhanced BEV features. Next, we detect the lane centerlines by regressing centerline points and identifying lane-related topological associations. Furthermore, we generate a mask for each centerline guided by the regressed centerline points. Lastly, we sample the mask and optimize them by incorporating the centerline points to improve the accuracy of centerlines and the reliability of topology.
• Figure 3: Qualitative evaluation of RoadPainter and comparison methods. Given multi-view images, RoadPainter achieves superior centerline detection performance compared to TopoNet in terms of completeness and accuracy. With the design of SD map interaction module, RoadPainter$^{*}$ precisely estimates the lane count at intersections.
• Figure 4: Qualitative analysis of instance mask of RoadPainter. The directly regressed centerlines from the transformer decoder, along with the instance masks and the refined centerlines resulting from the points-mask optimization module, are presented.
• Figure 5: Our method outperforms TopoNet in detecting high-curvature lanes due to the utilization of segmentation masks, which provide more precise geometric details.