MapGR: Learning Global Representation from Queries for Vectorized HD Map Construction

Comparison of a conventional DETR-like HD map decoder and MapGR's globally guided query decoder. — Conventional DETR-like map construction optimizes local queries independently. MapGR learns a map-level representation from the full query set and feeds that context back into decoding, improving global structure and producing smoother vector predictions.

Abstract

The online construction of vectorized high-definition (HD) maps is a cornerstone of modern autonomous driving systems. State-of-the-art approaches, particularly those based on the DETR framework, formulate this as an instance detection problem. However, their reliance on independent, learnable object queries results in a predominantly local query perspective, neglecting the inherent global representation within HD maps. In this work, we propose MapGR (Global Representation learning for HD Map construction), an architecture designed to learn and utilize a global representation from queries. Our method introduces two synergistic modules: a Global Representation Learning (GRL) module, which encourages the distribution of all queries to better align with the global map through a carefully designed holistic segmentation task, and a Global Representation Guidance (GRG) module, which endows each individual query with explicit, global-level contextual information to facilitate its optimization. Evaluations on the nuScenes and Argoverse 2 datasets validate the efficacy of our approach, demonstrating substantial improvements in mean average precision (mAP) compared with leading baselines.

Method

MapGR is a plug-in for DETR-like vector map decoders. It adds global supervision and context while leaving the underlying map encoder and prediction heads intact.

MapGR architecture with Global Representation Learning and Global Representation Guidance modules. — MapGR architecture. GRL and GRG are applied to the first two layers of the six-layer decoder.

Global Representation Learning (GRL)

GRL aggregates all local map queries into one global embedding and predicts a holistic rasterized map. Segmentation supervision aligns the full query distribution with the ground-truth map.

Global Representation Guidance (GRG)

GRG encodes the predicted map-level representation and fuses it into every local query, giving each query explicit scene-wide context during subsequent decoding.

Plug-and-play integration. The modules integrate with MapTR, MapTRv2, and MapQR. Applying them in the first two decoder layers provides the best balance between global distribution learning and final instance convergence.

Results

MapGR improves all three evaluated DETR-like baselines on nuScenes and Argoverse 2, covering lane dividers, pedestrian crossings, and road boundaries.

Baseline → **+ MapGR**. Gains are absolute mAP points. mAP₁ averages Chamfer AP over {0.2, 0.5, 1.0} m, mAP₂ over {0.5, 1.0, 1.5} m.
Dataset	Baseline	Epochs	mAP₁	mAP₂
nuScenes	MapTR	24	27.3 → 31.2 +3.9	50.3 → 54.5 +4.2
	MapTRv2	24	37.2 → 41.1 +3.9	61.5 → 65.0 +3.5
	MapQR	24	42.1 → 45.3 +3.2	65.3 → 68.1 +2.8
Argoverse 2	MapTR	6	33.7 → 35.3 +1.6	57.8 → 58.9 +1.1
	MapTRv2	6	38.4 → 39.6 +1.2	64.7 → 65.6 +0.9
	MapQR	6	42.1 → 42.9 +0.8	65.1 → 66.0 +0.9

With the longer 110-epoch schedule, MapGR improves MapTR by 4.9 mAP₂ on nuScenes.

Efficiency

Across all three baselines, MapGR retains nearly the same throughput. The paper reports approximately 1% runtime overhead (at most 1.2 ms) while adding 4-23% parameters, depending on the baseline.

Method	Params	FPS
MapTR	47.5M	24.2
+ MapGR	58.5M	23.8
MapTRv2	56.1M	19.6
+ MapGR	67.1M	19.4
MapQR	225.4M	17.9
+ MapGR	236.4M	17.7

Qualitative Results

Across six nuScenes examples, MapGR recovers more complete boundaries, parallel lanes, pedestrian crossings, and junction branches. Each row compares MapTRv2, MapQR, MapQR with MapGR, and ground truth.

Two nuScenes intersection examples comparing multiview images and maps predicted by MapTRv2, MapQR, MapQR with MapGR, and ground truth. — **Intersections and lane continuity.** MapGR preserves long, parallel lane structures and produces junction branches closer to the ground-truth topology.

Two nuScenes crossing and junction examples comparing multiview images and maps predicted by MapTRv2, MapQR, MapQR with MapGR, and ground truth. — **Crossings and multi-branch junctions.** Global guidance improves pedestrian-crossing alignment and reduces broken or duplicated road boundaries.

Two nuScenes curved-road examples comparing multiview images and maps predicted by MapTRv2, MapQR, MapQR with MapGR, and ground truth. — **Curved and enclosed road boundaries.** MapGR better follows long curves and enclosed islands while keeping lane dividers aligned with the scene geometry.

Analysis

Two line charts comparing query-coordinate mean absolute error across decoder transitions for MapQR and MapQR with MapGR. — **More stable refinement across decoder layers.** With MapGR, high-percentile coordinate changes fall sooner and remain lower after the first two stages, indicating a more stable query distribution in later decoding.

Baseline61.5

BSM62.1

Cross-attn.63.6

GIG (ours)65.0

Global feature generator. Appendix Table 9: GIG outperforms segmentation-mask and cross-attention alternatives for producing the global representation.

BibTeX

@inproceedings{qiu2026mapgr,
  title     = {Learning Global Representation from Queries for Vectorized {HD} Map Construction},
  author    = {Qiu, Shoumeng and Li, Xinrun and Long, Yang and Xue, Xiangyang and Ojha, Varun and Pu, Jian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}