AMap | CVPR 2026

Overview

Abstract

Online High-Definition (HD) map construction is pivotal for autonomous driving. While recent approaches leverage historical temporal fusion to improve performance, we identify a critical safety flaw in this paradigm: it is inherently spatially backward-looking. These methods predominantly enhance map reconstruction in traversed areas, offering minimal improvement for the unseen road ahead. Our analysis of downstream planning tasks reveals a severe asymmetry: while rearward perception errors are often tolerable, inaccuracies in the forward region directly precipitate hazardous driving maneuvers. To bridge this safety gap, we propose AMap, a novel framework for ahead-aware online HD mapping. We pioneer a distill-from-future paradigm, where a teacher model with privileged access to future temporal contexts guides a lightweight student model restricted to the current frame. This process implicitly compresses prospective knowledge into the student model, endowing it with look-ahead capabilities at zero inference-time cost. Technically, we introduce a Multi-Level BEV Distillation strategy with spatial masking and an Asymmetric Query Adaptation module to effectively transfer future-aware representations to the student's static queries. Extensive experiments on the nuScenes and Argoverse 2 benchmark demonstrate that AMap significantly enhances current-frame perception, especially in critical forward regions, while maintaining the efficiency of single current frame inference.

Method

Framework

Overview of the AMap framework. A future-aware teacher distills prospective knowledge into a current-frame student using multi-level BEV distillation and asymmetric query adaptation.

1. Distill-from-Future Paradigm

A teacher model sees future temporal context during training, while the student is restricted to the current frame and learns future-aware priors implicitly.

2. Multi-Level BEV Distillation

Multi-scale BEV features are distilled with spatial masking so supervision emphasizes useful spatial regions while reducing noisy transfer.

3. Asymmetric Query Adaptation

Bridges the gap between the teacher's richer representations and the student's static queries, making future-aware priors easier to absorb.

Benchmark

Quantitative Results

nuScenes validation set

Method	Temporal	Backbone	mAP	A-mAP	R-mAP	FPS
Temporal models
MapTR (GKT)	Yes	R50	51.28	52.23	55.22	14.7
MapTR (BEVFormer)	Yes	R50	53.29	53.69	55.18	15.6
StreamMapNet	Yes	R50	51.15	54.94	50.94	15.7
HRMapNet	Yes	R50	64.53	62.34	69.51	16.2
MapTracker	Yes	R18	69.34	66.86	76.51	22.3
MapTracker	Yes	R50	72.93	70.03	78.92	15.6

Current-frame models
MapTR	No	R50	44.10	45.01	48.87	17.2
GeMap	No	R50	51.26	51.61	54.74	13.7
MGMap	No	R50	53.15	55.13	56.78	12.4
MapTRv2	No	R50	61.38	64.21	61.82	14.4
MapQR	No	R50	66.35	68.38	65.59	12.4
MapTracker	No	R50	68.30	69.30	69.47	20.1
AMap	No	R50	69.26	70.19	69.61	20.1

MapTR	No	R18	32.35	36.34	32.76	31.4
MapTRv2	No	R18	57.15	60.81	56.17	16.6
MapQR	No	R18	62.31	64.70	63.13	13.7
MapTracker	No	R18	62.81	64.63	64.04	31.5
AMap	No	R18	64.49	66.28	65.11	31.5

Argoverse 2 validation set

Method	Temporal	Backbone	mAP	A-mAP	R-mAP
MapTracker	5 past	R50	76.87	73.00	83.82
MapTracker	5 future	R50	75.57	80.79	72.42
MapTracker	None	R18	63.69	65.29	66.10
AMap	None	R18	65.25	67.81	67.53

Visualization

Qualitative Results

Qualitative results on Argoverse 2 — **Argoverse 2.** Additional qualitative comparisons show better alignment and cleaner structural recovery in challenging scenes.

BibTeX

Citation

@article{li2025amap,
  title={AMap: Distilling Future Priors for Ahead-Aware Online HD Map Construction},
  author={Li, Ruikai and Li, Xinrun and Xie, Mengwei and Shan, Hao and Qiu, Shoumeng and Chang, Xinyuan and Fan, Yizhe and Xiong, Feng and Jiang, Han and Ren, Yilong and others},
  journal={arXiv preprint arXiv:2512.19150},
  year={2025}
}