PriorMaskMap: Robust Online Vectorized Map Construction with Biased Priors

Haoming Xu; Wei Li; Yu Hu

cs.CVOct 29, 2026

PriorMaskMap: Robust Online Vectorized Map Construction with Biased Priors

Authors: Haoming Xu, Wei Li, Yu Hu

Organizations: The Research Center for Intelligent Computing Systems, SKLP, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China · Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Hangzhou 310024, China · University of Chinese Academy of Sciences, Beijing 100049, China

Abstract

Online vectorized map construction provides autonomous vehicles with essential, real-time semantic and geometric scene understanding. While leveraging prior maps presents an opportunity to improve accuracy, they are often biased due to outdated information or temporary road changes. Existing methods, which typically treat priors as fully reliable or only consider limited noise patterns, suffer severe performance degradation when these priors are imperfect. To address this, we propose PriorMaskMap, a robust framework for prior-guided vectorized map construction. Its first innovation is a Prior-Aware Confidence Estimator (PACE) that evaluates the input rasterized map against current visual observations. It assigns confidence scores to regions classified as reliable, erroneous, or blank, creating a foundational mask that guides two subsequent branches. In one branch, the Dual-Prior Query Generator (DPQG) selectively filters and embeds both rasterized and vectorized priors to initialize informative queries from trustworthy map structures. In the other, the Confidence-Guided Prior Fusion (CGPF) module spatially propagates and refines the mask, using the refined version to govern the fusion of BEV features with the prior map, thereby providing rich contextual representations for the map decoder. Evaluation on the nuScenes dataset shows that PriorMaskMap achieves superior mapping precision over existing approaches and exhibits marked robustness when provided with priors containing apparent errors. Our code is released at https://github.com/healenrens/PriorMaskMap.

Figures & tables

Figure 1: Overall framework of PriorMaskMap. Given multi - view images and a biased prior map, PACE predicts grid - wise confidence over the rasterized prior. DPQG converts confident vector and raster structures into prior - aware queries, while CGPF diffuses and refines the confidence field to produce a BEV map context for the decoder.

Panel of Figure 2 — Figure 2: (a) Architecture of the Confidence-Guided Prior Fusion (CGPF) module. It contains a Spatial Confidence Propagation (SCP) block that produces a diffused confidence mask and a Confidence Refinement Module (CRM) that refines prior weights and outputs the BEV map context. (b) Illustration of the binarized refined confidence mask. For clarity, it is presented as a binary map where white indicates reliable regions of the prior and black indicates erroneous ones. Compared to the Ground Truth (GT), erroneous regions in the Biased Prior (red dashed circles) are excluded as black in the map, while reliable regions (green dashed circles) are preserved as white.

Methods	Backbone	Epoch	Prior	Err. ∗	$AP_{\text{ped}}$	$AP_{\text{div}}$	$AP_{\text{boun}}$	mAP
StreamMapNet [ 39 ] WACV24	R50	30	–	–	56.9	55.9	61.4	58.1
MGMap [ 19 ] CVPR24	R50	24	–	–	61.8	65.0	67.5	64.8
HIMap [ 43 ] CVPR24	R50	30	–	–	62.6	68.4	69.1	66.7
MapQR [ 21 ] ECCV24	R50	24	–	–	63.4	68.0	67.7	66.4
Mask2Map [ 4 ] ECCV24	R50	24	–	–	70.6	71.3	72.9	71.6
MapTRv2 [ 16 ] IJCV25	R50	24	–	–	59.8	62.4	62.4	61.5

Table 1: Comparison with state - of - the - art methods on the nuScenes dataset. The first block lists methods that do not rely on priors or history for map construction, while the second block lists methods that leverage various priors. The third block lists models retrained and evaluated only with biased priors under our mixed-bias setting. ∗ Err. denotes that, during evaluation, only biased prior maps are used as inputs.

\dagger

The results are trained with our mixed data from scratch by replacing history or prior maps.

Figure 3: Visualization comparing our proposed model against the HRMapNet † .

Method	Biased Type	Epoch	$AP_{\text{ped}}$	$AP_{\text{div}}$	$AP_{\text{boun}}$	mAP
SQD-MapNet †	Addition	24	55.4	53.3	57.9	55.5
	Reduction		55.4	53.3	57.9	55.5
	Widening		55.4	53.4	57.9	55.6
	Narrowing		55.4	53.3	57.9	55.5
PrevPredMap †	Addition	24	63.1	60.6	65.9	63.2
	Reduction		63.1	60.6	65.9	63.2

Table 2: Performance comparison on the nuScenes dataset with mixed bias prior. † as in Tab. 1 and the first two results are not typos but follow from their respective processing mechanisms. Bold numbers indicate the best mAP for each biased type.

Method	StreamMapNet				Geo-split
Method	$AP_{\text{ped}}$	$AP_{\text{div}}$	$AP_{\text{boun}}$	mAP	$AP_{\text{ped}}$	$AP_{\text{div}}$	$AP_{\text{boun}}$	mAP
HRMapNet †	38.0	70.0	71.3	59.7	51.9	65.7	64.7	60.8
Ours	40.2	78.7	74.8	64.6 (+4.9)	49.9	75.5	67.3	64.2 (+3.4)

Table 3: Quantitative comparison with the baseline on StreamMapNet and Geo-split datasets. † as in Tab. 1 . Best results are highlighted in bold . Both models are trained for 24 epochs.

Vectorized Prior	SCP	CRM	$AP_{\text{ped}}$	$AP_{\text{div}}$	$AP_{\text{boun}}$	mAP
✗	✗	✗	74.5	69.3	70.3	71.4
✓	✗	✗	79.1	74.4	73.9	75.8 (+4.4)
✓	✗	✓	80.2	73.5	75.9	76.5
✓	✓	✗	79.7	74.9	74.5	76.4
✗	✓	✓	80.5	72.6	71.6	74.8 (+3.4)
✓	✓	✓	79.6	75.3	75.1	76.7 (+5.3)

Table 4: Ablation study of DPQG, SCP, and CRM.

HD Map Construction

High-definition map construction turns sensor and geographic data into precise, structured road maps for autonomous driving and simulation.

See all 41 papers

MapLightning: Online Vectorized HD Map Construction with 1D Map Tokens
Oct 1, 2026Shen Zheng, Anurag Ghosh, Mani Ramanagopal +1HD Map Construction
Carnegie Mellon University
Unified Map Prior Encoder for Mapping and Planning
May 4, 2026Zongzheng Zhang, Sizhe Zou, Guantian Zheng +12End-to-End Autonomous Driving Autonomous Driving
Institute for AI Industry Research (AIR), Tsinghua University. · Bosch Corporate Research, China. · Shanghai Jiao Tong University. +1
OptiMVMap: Offline Vectorized Map Construction via Optimal Multi-vehicle Perspectives
Apr 18, 2026Zedong Dan, Zijie Wang, Wei Zhang +6Multi-View Learning Autonomous Driving
Sun Yat-sen University · Zhongguancun Academy · Shenzhen Loop Area Institute +2

Oct 1, 2026J/K move · Enter open · S save

MapLightning: Online Vectorized HD Map Construction with 1D Map Tokens

Shen Zheng, Anurag Ghosh, Mani Ramanagopal, Srinivasa Narasimhan

Carnegie Mellon University

HD Map Construction

Open paper AskPDF arXiv

Online vectorized HD map construction is essential for scaling safe autonomous driving and requires accurate, real-time inference. Prior methods typically rely on dense bird's-eye-view (BEV) grids as the intermediate representation. We propose MapLightning, which replaces the dense BEV grid with a compact set of 1D learnable map tokens. To construct map tokens from image features, we choose self-attention over vanilla cross-attention because it enables joint interactions and contextual aggregation among image and map tokens. Our transformer-based mapper concatenates map and image tokens, applies full self-attention, discards the image tokens, and retains the updated map tokens for decoding. This design offers three advantages. First, our representation is efficient, using fewer tokens, consuming less memory, and running faster. Second, the lightweight design allows the map decoder to use full rather than deformable cross-attention for better global context. Third, unlike BEV-based methods, our network does not use camera projection parameters, making it robust to camera-extrinsic perturbations. MapLightning uses up to 16.7

\times

fewer intermediate tokens than dense BEV-based methods and achieves state-of-the-art accuracy and efficiency on nuScenes and Argoverse~~2. Its lightweight variant surpasses MapTRv2 by +10.1 mAP on nuScenes and +16.2 mAP on Argoverse~~2, while delivering 1.73

\times

faster inference (40+ FPS) with 53% less memory. We further show improvements on uncertainty-aware map construction and downstream trajectory prediction. Code and models will be released.

Abstract

HD Map Construction

Explore similar work

MapLightning: Online Vectorized HD Map Construction with 1D Map Tokens

Unified Map Prior Encoder for Mapping and Planning

OptiMVMap: Offline Vectorized Map Construction via Optimal Multi-vehicle Perspectives