.png){kind=icon}
Return
Axis-Selective Query Initialization for Online Vectorized HD Map Construction
.png){kind=icon}
PRE
AI
DOI：10.1109/tvt.2026.3749615.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Online vectorized high-definition (HD) map construction aims to predict local road elements from onboard sensor data. Recent DETR-based methods usually represent road elements with randomly initialized queries and associated initial point coordinates that describe the geometry of each map element. However, road elements in BEV space are sparse, elongated, and irregular, making fixed-distribution initial point coordinates poorly aligned with their geometries. Meanwhile, random queries are input-agnostic learnable embeddings and lack semantic cues from the current scene. To better align initial queries with road geometry and input features, we propose MapAS. As the core of MapAS, Axis-Selective Query Initialization (ASQI) exploits the axis-wise geometric prior of road elements to generate input-dependent queries from BEV features. Specifically, ASQI uses a mask to highlight BEV features related to road elements, compresses these features along the row and column axes of the BEV feature map, and selects high-quality representative axis-wise features. To further enhance categoryspecific information during axis-wise compression, we introduce a Class-Aware Mask Generator (CAMG) to add category priors into query generation. Moreover, input-dependent queries provide suitable candidates for a Hybrid-Layer Scheme (HLS), a layer-wise hybrid matching scheme, without adding extra queries. Experiments on nuScenes and Argoverse2 demonstrate the effectiveness of MapAS across diverse mapping scenarios, including standard, long-range, and 3D mapping settings, while maintaining efficient inference.
Keywords:
vectorized HD map construction
online mapping
query initialization
autonomous driving
Journal
IF:
7.1
Papers:
1.8W
Citations:
6.6W
Organization
Cited Papers
No cited papers available