[Paper Review] Online Map Vectorization for Autonomous Driving: A Rasterization Perspective

Gongjie Zhang, Jiahao Lin+6 more|arXiv (Cornell University)|Jun 18, 2023

Geographic Information Systems Studies16 citations

TL;DR

The paper introduces MapVR, a differentiable rasterization framework that enhances vectorized HD map predictions by supervising vector outputs with rasterized HD maps, plus a more sensitive rasterization-based evaluation metric.

ABSTRACT

Vectorized high-definition (HD) map is essential for autonomous driving, providing detailed and precise environmental information for advanced perception and planning. However, current map vectorization methods often exhibit deviations, and the existing evaluation metric for map vectorization lacks sufficient sensitivity to detect these deviations. To address these limitations, we propose integrating the philosophy of rasterization into map vectorization. Specifically, we introduce a new rasterization-based evaluation metric, which has superior sensitivity and is better suited to real-world autonomous driving scenarios. Furthermore, we propose MapVR (Map Vectorization via Rasterization), a novel framework that applies differentiable rasterization to vectorized outputs and then performs precise and geometry-aware supervision on rasterized HD maps. Notably, MapVR designs tailored rasterization strategies for various geometric shapes, enabling effective adaptation to a wide range of map elements. Experiments show that incorporating rasterization into map vectorization greatly enhances performance with no extra computational cost during inference, leading to more accurate map perception and ultimately promoting safer autonomous driving.

Motivation & Objective

• Motivate precise online HD map construction for autonomous driving and address limitations of existing vectorization and evaluation methods.
• Introduce a rasterization-based evaluation metric that is more sensitive to geometric deviations than Chamfer distance.
• Propose MapVR, a framework that bridges vectorized outputs with rasterized HD maps through differentiable rasterization for geometry-aware supervision.

Proposed method

• Define a differentiable rasterizer that renders line elements via an exponential distance-based mask and polygon elements via a signed distance with a sigmoid to produce HD masks.
• Apply rasterized HD map supervision (dice loss) to guide vectorized predictions, removing reliance on equidistant point targets.
• Incorporate a direction regularization loss to encourage smooth, curvature-aware vector outputs.
• Maintain inference-time efficiency by discarding the rasterization module after training.
• Demonstrate compatibility with existing map vectorization architectures (e.g., MapTR) as a plug-in.

Experimental results

Research questions

• RQ1 How does rasterization-based supervision affect precision and robustness of vectorized HD map elements in autonomous driving scenarios?
• RQ2 Does a rasterization-based evaluation metric better detect small geometric deviations than Chamfer-distance-based metrics?
• RQ3 Can MapVR improve state-of-the-art map vectorization (e.g., MapTR) without increasing inference cost?
• RQ4 How do rasterization resolution, softness, and geometry-aware rendering impact learning and performance for different map element shapes (lines vs polygons)?

Key findings

• MapVR improves map vectorization accuracy across multiple datasets and settings, notably with raster-based evaluation (AP_raster) showing larger gains than Chamfer-based AP.
• On nuScenes Map (basic), MapVR yields a notable improvement over MapTR, with larger margins when using multi-modality inputs.
• On nuScenes Map (extended) and Argoverse2, MapVR achieves state-of-the-art performance under both metrics, validating robustness across datasets.
• 6V-mini-v0.4 results show MapVR consistently boosts elements like lanes and crosswalks, demonstrating effectiveness in complex real-world scenes.
• A rasterization-based evaluation metric provides more sensitive and geometry-informed matching than Chamfer distance, aligning better with driving safety requirements.