FastMap: Fast Queries Initialization Based Vectorized HD Map Reconstruction Framework (2503.05492v1)

Abstract: Reconstruction of high-definition maps is a crucial task in perceiving the autonomous driving environment, as its accuracy directly impacts the reliability of prediction and planning capabilities in downstream modules. Current vectorized map reconstruction methods based on the DETR framework encounter limitations due to the redundancy in the decoder structure, necessitating the stacking of six decoder layers to maintain performance, which significantly hampers computational efficiency. To tackle this issue, we introduce FastMap, an innovative framework designed to reduce decoder redundancy in existing approaches. FastMap optimizes the decoder architecture by employing a single-layer, two-stage transformer that achieves multilevel representation capabilities. Our framework eliminates the conventional practice of randomly initializing queries and instead incorporates a heatmap-guided query generation module during the decoding phase, which effectively maps image features into structured query vectors using learnable positional encoding. Additionally, we propose a geometry-constrained point-to-line loss mechanism for FastMap, which adeptly addresses the challenge of distinguishing highly homogeneous features that often arise in traditional point-to-point loss computations. Extensive experiments demonstrate that FastMap achieves state-of-the-art performance in both nuScenes and Argoverse2 datasets, with its decoder operating 3.2 faster than the baseline. Code and more demos are available at https://github.com/hht1996ok/FastMap.

Overview of FastMap: Fast Queries Initialization Based Vectorized HD Map Reconstruction Framework

The paper presents FastMap, a novel approach to vectorized high-definition (HD) map reconstruction specifically designed for improving the computational efficiency of algorithms used in autonomous driving systems. This research addresses the critical issue of decoder redundancy inherent in existing frameworks. By optimizing the decoder structure and leveraging efficient query initialization strategies, FastMap presents a compelling advancement in online HD map construction.

Core Innovations

The authors identify inefficiencies in current vectorized map reconstruction methods, such as those based on the DEtection TRansformers (DETR) framework, which require the stacking of multiple decoder layers to maintain performance. This method significantly increases the computational demands and negatively impacts real-time processing capabilities. FastMap provides an alternative through a novel decoder architecture that employs only a single-layer, two-stage transformer that retains multilevel representation.

Key innovations introduced in FastMap include:

• Heatmap-Guided Query Generation: Rather than using random query initializations, FastMap utilizes a heatmap-guided module to generate queries during decoding. This strategic approach not only maps image features efficiently into structured vectors but also incorporates learnable positional encodings to enhance decoder precision.
• Geometry-Constrained Loss Mechanism: To differentiate highly homogeneous features often found in traditional point-to-point loss calculations, FastMap implements a point-to-line loss mechanism. This enables the model to focus on minimizing distances from predicted points to ground truth curves, which represents a significant departure from previous methods and reduces forced alignment of prediction points with annotated points.

Experimental Validation

The FastMap framework exhibits state-of-the-art performance in empirical evaluations, outperforming existing models on benchmark datasets like nuScenes and Argoverse2. Noteworthy results include:

• The decoder operates 3.2 times faster than baseline state-of-the-art models, showcasing significant improvements in computational efficiency while maintaining high accuracy in map reconstructions.
• FastMap-large version achieves an impressive 68.1 mAP on the nuScenes dataset, thereby representing a substantial advancement over existing methods.

Implications and Future Work

The implications of this research are multifaceted, affecting both the theoretical architecture of HD map construction frameworks and their practical deployment in autonomous driving systems. The streamlined, efficient decoder design championed by FastMap could pave the way for broader adoption of vectorized maps in resource-constrained environments, such as embedded systems in vehicles.

From a theoretical perspective, FastMap's innovations can drive the development of future mapping models that employ more sophisticated attention mechanisms and more nuanced loss functions to improve precision without increasing computational load.

Potential future directions include extending the heatmap-guided query generation module to support additional sensor modalities or exploring iterative refinement processes that further lessen the burden on onboard computational resources.

Overall, the FastMap framework represents a highly effective model for addressing the specific challenges faced in real-time map reconstruction for autonomous vehicles, providing significant contributions to the field with its enhanced speed and precision.