AMap: Distilling Future Priors for Ahead-Aware Online HD Map Construction

TL;DR

The paper challenges the prevailing backward-looking paradigm in online HD map construction by showing that forward perception is critical for safe planning. It introduces AMap, a distill-from-future framework where a teacher with future context guides a current-frame student through BEV feature distillation with spatial masking and asymmetric query transfer, yielding ahead-aware representations without extra inference cost. The authors propose novel metrics A-mAP and R-mAP to separately evaluate forward and rear map quality and demonstrate substantial gains in the forward region on nuScenes and Argoverse 2, outperforming some temporal models while preserving single-frame efficiency. This work offers a practical, plug-and-play training paradigm that realigns map perception with planning priorities, potentially enhancing the safety and reliability of autonomous driving systems.

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. Crucially, 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. Most notably, it outperforms state-of-the-art temporal models in critical forward regions while maintaining the efficiency of single current frame inference.

Figures (11)

• Figure 1: (a) Performance Bias in SOTA Temporal Methods. Current state-of-the-art temporal mapping methods exhibit a severely weeker mAP in front region (A-mAP) than it in rear views (R-mAP), as illustrated in the orange-shaded region. (b) & (c) Back-Aware Pipeline vs. Our Ahead-aware Pipeline. While the (b) Back-Aware mapping paradigm enhances perception in the vehicle's rearward area, this comes at the cost of limited gains in forward spatial awareness. In contrast, our paradigm (c) effectively achieves forward perception gains by acquiring future priors through knowledge distillation.
• Figure 2: Impact of directional masking the outputs of the online HD mapper on downstream trajectory prediction. As the Forward Mask ratio increases (red line), metrics such as minADE, minFDE, and MR deteriorate significantly, whereas the Backward Mask (blue line) has negligible impact. This empirical evidence demonstrates that downstream tasks are critically dependent on the ahead-aware mapping outputs.
• Figure 3: The overview of our proposed AMap framework.
• Figure 4: BEV feature visualization comparison among: (b) historical sequence-based MapTracker, (c) our teacher model with future priors, (d) the baseline student model, and (e) the student model with our proposed distillation strategy. Blue boxes and red boxes indicate the advantages of historical sequences and future sequences, respectively.
• Figure 5: Qualitative results of the baseline student model (a), AMap (Ours) (b), and Ground-Truth (c).