Hi-D Maps: High-Definition Mapping

• Hi-D Maps are high-definition maps that provide detailed lane-level geometry and rich semantics for precise localization and perception.
• They fuse sensor inputs like cameras and LiDAR using vectorized and rasterized approaches to create accurate, machine-readable representations.
• Neural construction and temporal fusion methods enhance global map consistency, enabling robust autonomous driving performance.

Searching arXiv for the cited works to ground the article in current literature. In contemporary autonomous-driving research, Hi-D Maps are most commonly understood as high-definition maps: machine-readable, 1:1 geometric and semantic models of the road environment that provide lane-level geometric accuracy, rich semantics, and topological rules for localization, perception, prediction, and planning (Wijaya et al., 2024). In the recent learning-based literature, they are typically represented in bird’s-eye view as vectorized instances such as lane dividers, road boundaries, and pedestrian crossings, inferred from cameras, LiDAR, or top-down tiles and, in some systems, fused across multiple traversals into world-coordinate global maps (Fan et al., 2024).

1. Definition, layers, and functional role

HD maps encode the static road environment at a granularity substantially beyond consumer navigation maps. The reviewed literature describes lane-level geometric accuracy, often at the 10–20 cm scale, alongside rich semantics such as lane boundaries, centerlines, crosswalks, stop lines, traffic signs, traffic lights, poles, barriers, curbs, sidewalks, and connectivity or rule layers including turn restrictions and speed limits (Wijaya et al., 2024). A related survey organizes this content into road, lane, and localization layers, emphasizing that HD maps encode not only visible road geometry but also structural and semantic priors that may be occluded in current sensor observations (Bao et al., 2022).

These maps function as a static, extended sensor. In the autonomous-driving stack they support high-precision localization, constrain trajectory generation, provide priors for perception, and supply lane-graph structure for planning and prediction (Wijaya et al., 2024). This division of labor is reflected in map standards and production systems. Lanelet2 decomposes map content into physical objects such as points and line strings, relational objects such as lanelets and areas, and topological relations such as successor and adjacent links, whereas OpenDRIVE centers representation on a road reference line, lane sections, and road objects (Bao et al., 2022).

A recurrent distinction in the literature is between local online map perception and production-grade global mapping. Recent online methods infer only ego-centric neighborhoods, whereas industrial workflows require complete, semantically correct, globally consistent maps expressed in a world coordinate frame (Fan et al., 2024). This distinction is central to nearly all recent work on neural HD-map construction.

2. Geometric and semantic representations

The dominant learned representation is vectorized. Static map elements are encoded as polylines or polygons in BEV, usually for categories such as pedestrian crossings, lane dividers, and road boundaries (Fan et al., 2024). In HIMap, HD map construction is formulated as direct set prediction of vectorized elements, where each instance has a class $c_i$, an ordered sequence of $P$ BEV points $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$, and an element mask $M_i \in [0,1]^{H \times W}$ (Zhou et al., 2024). That formulation makes explicit a distinction between point-level information, which captures local coordinates and tangent structure, and element-level information, which captures global shape and semantics.

Rasterized representations remain important as training targets, intermediate states, or fusion domains. GNMap rasterizes each vectorized tile to a grayscale image during pretraining,

$\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$

and predicts a multi-channel semantic map during finetuning,

$\mathcal{Y} \in \mathbb{R}^{h \times w \times c},$

with background, pedestrian crossing, lane divider, and road boundary channels (Fan et al., 2024). PolyMap pushes this raster emphasis further by reinterpreting road elements as rasterized polygons, using instance segmentation to predict masks and a Potrace-based module to recover vectorized map elements (Gao et al., 8 Nov 2025). This design is motivated by the observation that fixed-length point sequences impose an awkward parameterization on variable-length road structures.

A second major representational axis is the relation between perspective-view and BEV features. HybriMap argues that sequential PV$\rightarrow$BEV pipelines lose early-stage perspective cues and therefore introduces “hybrid clues” from both PV and BEV features (Zhang et al., 2024). HeightMapNet, in turn, treats road-surface height as an explicit latent variable and computes BEV features through a height-weighted aggregation,

$$F_i^{bev} = \mathbb{AP}\big( Height^{pro} \odot Height_i^{range} \big),$$

rather than assuming a flat or implicitly handled ground plane (Qiu et al., 2024). Across these systems, the representational trend is toward coupling vectorized outputs with richer dense intermediate structure rather than treating polyline regression as an isolated endpoint.

3. Construction paradigms

Classical HD-map generation begins with mobile mapping systems combining LiDAR, cameras, GNSS, and IMU, followed by preprocessing, SLAM or registration, semantic extraction, vectorization, and manual verification (Bao et al., 2022). Surveys highlight NDT, ICP, LOAM, LIO-SAM, FAST-LIO2, and related graph-based or filter-based pipelines as the geometric backbone of offline mapping, especially for globally consistent point-cloud maps and subsequent semantic annotation (Wijaya et al., 2024). These approaches remain strong in geometric fidelity but are labor-intensive, costly, and slow to update at scale.

Neural construction methods shift the bottleneck from manual vectorization to learned BEV perception. HIMap is representative of direct vector-set prediction: it introduces a hybrid representation,

$$\mathbf{Q}^{h} \in \mathbb{R}^{E \times (P+1) \times C},$$

called HIQuery, decomposed into point queries and element queries, and refines them through a point-element interactor and a point-element consistency constraint (Zhou et al., 2024). On nuScenes, HIMap reports $77.8$ mAP with multi-view RGB + LiDAR at 110 epochs, surpassing previous state of the art by at least $P$0 mAP (Zhou et al., 2024). The significance is not only numerical; the method formalizes the claim that point-level regression alone is insufficient for preserving global shape and avoiding element entanglement.

HybriMap addresses a different bottleneck: information loss during PV$P$1BEV transformation. Its Dual Enhancement Module explicitly integrates projected BEV enhancement cues into perspective heatmaps and implicitly refines BEV content features, yielding $P$2 mAP on nuScenes under the general threshold set and $P$3 mAP on Argoverse 2, both above strong MapTRv2 baselines in the reported settings (Zhang et al., 2024). HeightMapNet targets yet another failure mode, namely poor handling of road height, irrelevant background, and single-scale BEV reasoning. With explicit height modeling, foreground-background separation, and multi-scale BEV fusion, HeightMapNet$P$4 reports $P$5 general-threshold mAP and $P$6 tighter-threshold mAP on the nuScenes test set, and $P$7 mAP on Argoverse 2 validation (Qiu et al., 2024). Taken together, these works indicate that “HD-map learning” is not a single architectural recipe but a sequence of attempts to repair specific information bottlenecks in BEV map construction.

4. From local perception to global and temporally consistent maps

A persistent misconception in the online mapping literature is that accurate local perception is already equivalent to global HD-map construction. GNMap directly contradicts this. It observes that even strong online methods such as MapTR and PivotNet typically recover only about $P$8 of the true local HD map elements around the vehicle in one tour, leaving the problem of how to fuse multiple incomplete, inconsistent local tiles into a world-coordinate global map (Fan et al., 2024). GNMap formulates tile fusion as

$P$9

implemented by a multi-layer, attention-based autoencoder pretrained for masked map completion and finetuned for multi-tour semantic fusion (Fan et al., 2024). On a real-world Mainland China dataset, GNMap reports $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$0, $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$1, and $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$2, outperforming MapTR’s $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$3 F1 by more than 5 points and a GMM fusion baseline’s $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$4 F1 by more than 10 points; pretraining adds roughly 8–9 F1 points compared with training from scratch (Fan et al., 2024). The paper further states that GNMap has been deployed at Navinfo Co., Ltd. and reaches “the level of industrial usage with a small amount of manual modification” (Fan et al., 2024).

Temporal consistency is a second globalizing pressure. HisTrackMap argues that query propagation alone does not preserve sufficiently explicit geometry over time and introduces instance-level historical rasterization maps,

$\mathbf{P}_i \in \mathbb{R}^{P \times 2}$5

one per tracked global instance, together with Map-Trajectory Prior Fusion and a new global metric, G-mAP, for sequence-level geometry (Yang et al., 10 Mar 2025). On nuScenes validation, HisTrackMap improves over a reproduced MapTracker baseline at 24 epochs from $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$6 to $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$7 mAP, from $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$8 to $\mathbf{P}_i \in \mathbb{R}^{P \times 2}$9 C-mAP, and from $M_i \in [0,1]^{H \times W}$0 to $M_i \in [0,1]^{H \times W}$1 G-mAP; at 72 epochs it reports $M_i \in [0,1]^{H \times W}$2 mAP and $M_i \in [0,1]^{H \times W}$3 G-mAP (Yang et al., 10 Mar 2025). PredMapNet extends the same trajectory of ideas by combining a Semantic-Aware Query Generator, a History Rasterized Map Memory, a History-Map Guidance module, and a Short-Term Future Guidance module. Its ablations show best performance with four history frames, reaching $M_i \in [0,1]^{H \times W}$4 mAP and $M_i \in [0,1]^{H \times W}$5 C-mAP in the reported nuScenes-old-split setting (Lang et al., 18 Feb 2026). This suggests that explicit history and short-term prediction are becoming structural components of online HD-map construction rather than optional refinements.

A different input regime appears in TileTracker, which replaces perspective cameras with top-down tile images derived from LiDAR. It treats mapping as tracking-based vector HD mapping over 30 m $M_i \in [0,1]^{H \times W}$6 60 m top-down tiles and reports that TileTracker consistently outperforms a StreamTileNet baseline; for divider lines, intensity tiles yield mean AP $M_i \in [0,1]^{H \times W}$7 versus $M_i \in [0,1]^{H \times W}$8 for color tiles (Mahdavian et al., 2024). The methodological implication is that BEV-native inputs can simplify the front end while retaining the temporal advantages of instance tracking.

5. Evaluation, production workflows, and map maintenance

Evaluation has become increasingly stratified. Single-frame vector-map work usually reports mean Average Precision using Chamfer-distance thresholds such as $M_i \in [0,1]^{H \times W}$9 m or $\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$0 m, averaged over classes like pedestrian crossings, lane dividers, and road boundaries (Zhou et al., 2024). GNMap instead evaluates pixel-level precision and recall under a 0.5 m Euclidean tolerance and summarizes fusion quality with

$\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$1

reflecting its tile-fusion formulation (Fan et al., 2024). HisTrackMap’s G-mAP adds a distinct sequence-level perspective by measuring the geometric quality of globally merged predictions in a world frame (Yang et al., 10 Mar 2025). The broader review literature correspondingly separates spatial, semantic, and temporal map quality dimensions, including geometric accuracy, semantic completeness, classification accuracy, update frequency, and temporal consistency (Wijaya et al., 2024).

Production workflows remain hybrid. A typical industrial pipeline described in the literature is: repeated data collection by survey vehicles; online local tile generation by onboard neural models; offline fusion into global HD tiles; and manual QA or editing for residual errors and safety compliance (Fan et al., 2024). This division explains why “online HD map learning” and “HD map production” should not be conflated. Even where learned systems are strong, manual verification remains part of the pipeline, and large-scale maintenance still depends on change detection, update scheduling, and sometimes crowdsourced observations (Wijaya et al., 2024).

Map maintenance has therefore emerged as a field in its own right. Reviews distinguish single-session and multi-session change detection, direct and incremental update strategies, and map-heavy versus map-lite operating regimes (Wijaya et al., 2024). A plausible implication is that future HD-map systems will be judged not only by instantaneous mAP but by the stability of their global geometry, their update latency, and their compatibility with collaborative or fleet-scale maintenance loops.

6. Other technical meanings of “Hi-D maps”

The label is not exclusive to autonomous driving. In visualization research, “Hi-d maps” denotes an interactive visualization technique for multi-dimensional categorical data that maps the full data space to a 2D regular polygonal region and recursively cuts that polygon with lines parallel to dimension-associated sides (Reza et al., 10 Jul 2025). The method uses orientation, hue, lightness, saturation, line thickness, countable glyphs, and text to encode categorical intersections and their sizes, and supports dimension reordering and hierarchical browsing (Reza et al., 10 Jul 2025). Here, “Hi-D” refers to high-dimensional categorical visualization rather than high-definition road maps.

In astronomy and cosmology, closely related labels are used for HI-based maps. One extragalactic study constructs global HI-deficiency maps by computing

$\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$2

and then mapping the sky distribution of HI-deficient and HI-excess galaxies relative to optical scaling relations (Dénes et al., 2014). Another paper develops all-sky maps of the number of HI clouds along the line of sight by Gaussian-decomposing HI spectra and counting distinct kinematic components per HEALPix superpixel (Panopoulou et al., 2020). A further line of work uses generative models to synthesize 3D HI intensity maps; “Large, fast and accurate HI intensity maps with latent overlap diffusion” couples a halo-field predictor with a conditional variational diffusion model to generate $\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$3 21 cm brightness-temperature cubes, reporting power-spectrum accuracy within $\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$4 for $\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$5 after training on $\mathcal{X} \in \mathbb{R}^{h \times w \times 1},$6 hydrodynamical simulations (Mishra et al., 9 Jun 2025).

These usages are technically unrelated, but they show that “Hi-D Maps” is context-dependent. In current autonomous-driving research it primarily designates high-definition map construction and maintenance; in other literatures it names either high-dimensional visualization or neutral-hydrogen-derived map products. Within autonomous driving itself, the main trajectory of the field runs from classical SLAM-backed mobile mapping to end-to-end vectorized BEV perception, and from per-frame local inference toward explicit multi-tour fusion, instance-level temporal memory, and global geometry assessment (Bao et al., 2022).