EGC-VMAP: End-to-End City-Scale HD Map Fusion

• The paper introduces EGC-VMAP, which fuses multi-vehicle, multi-temporal vectorized maps using a Trip-Aware Transformer and hierarchical matching.
• It integrates on-vehicle lightweight mapping with cloud-side aggregation to overcome limitations of LiDAR surveys and single-vehicle perception.
• The architecture achieves significant mAP gains under challenging conditions while reducing manual annotation costs by approximately 90% at city scale.

Searching arXiv for the specified paper to ground the article in the source record. {"query":"(Feng et al., 11 Jul 2025) End-to-End Generation of City-Scale Vectorized Maps by Crowdsourced Vehicles","max_results":5} EGC-VMAP, short for End-to-End Generation of City-Scale Vectorized Maps by Crowdsourced Vehicles, is a framework for generating high-precision, city-scale vectorized maps by aggregating asynchronous, heterogeneous vectorized map elements collected by many standard vehicles over time. It is positioned against two established paradigms: offline LiDAR surveys, which provide very high accuracy but require specialized survey vehicles, expensive high-precision sensors, and labor-intensive manual labeling, and single-vehicle perception methods, which can work onboard standard vehicles but suffer from limited field-of-view, sensor noise, occlusions, and adverse weather or lighting. EGC-VMAP addresses these limitations through edge-computing of per-vehicle vectorized maps, cloud-side aggregation, a Trip-Aware Transformer, hierarchical matching during training, and a multi-objective loss that jointly optimizes classification, point accuracy, edge direction, and auxiliary segmentation (Feng et al., 11 Jul 2025).

1. Problem setting and design objective

The framework targets the creation of unified, high-precision HD maps from crowdsourced vehicles rather than from specialized survey fleets. In the formulation presented for EGC-VMAP, each vehicle first produces a lightweight vectorized map onboard, and these per-vehicle outputs are subsequently fused in the cloud. This shifts the operating point away from transmitting raw sensor data and toward transmitting structured map elements.

The motivating deficiencies of traditional pipelines are explicit. Offline LiDAR surveys are described as costly and slow because production depends on specialized hardware and manual labeling. Single-vehicle methods such as HDMapNet and MapTR react faster and run on standard vehicles, but the resulting vectorized maps can contain missing elements, geometric inaccuracies, and structural inconsistencies. The stated objective of EGC-VMAP is therefore to fuse multi-vehicle, multi-temporal perceived map elements into a city-scale HD map with improved semantic consistency and geometric robustness (Feng et al., 11 Jul 2025).

The framework is also defined as end-to-end. In context, this means that the multi-trip fusion process, the decoding of vectorized map elements, and the supervision strategy are integrated into a unified learning system rather than implemented as a brittle multi-stage pipeline. A plausible implication is that the architecture is intended not merely to aggregate redundant observations, but to learn which cross-trip correspondences are geometrically and semantically coherent.

2. System architecture

The pipeline begins with crowdsourced maps $X$, followed by arc-length resampling and normalization, then a Trip-Aware Transformer, a hierarchical query decoder, and finally a predicted vector map $\hat{Y}$. During training, hierarchical matching and multi-task supervision are applied (Feng et al., 11 Jul 2025).

Input representation

Each perceived map $E_k$, where $k = 0 \ldots N_e - 1$, consists of $N_v$ vector elements $V_m = \{P_n, c\}_{n=0}^{N_p-1}$. From these elements, the framework extracts:

• $$X_{\text{coord}} \in \mathbb{R}^{n_e \times n_v \times n_p \times 2}$$
• $$X_{\text{cls}} \in \mathbb{R}^{n_e \times n_v \times n_p \times 1}$$

Arc-length-based resampling is used to equalize $N_p$ across elements, and $(x,y)$ coordinates are min-max normalized globally. The same preprocessing is used during inference.

Trip-Aware Transformer

The encoder incorporates a semantic MLP that encodes per-point one-hot class into $\hat{Y}$0-dimensional features. Positional encoding learns both a trip ID embedding and an element position embedding. Semantic and positional features are concatenated and passed through an MLP to obtain $\hat{Y}$1-dimensional representations. The encoder then applies $\hat{Y}$2 layers of standard self-attention and feed-forward blocks to propagate information across trips, elements, and points.

The decoder is hierarchical and is described as inspired by MapTR. It uses:

• instance queries $\hat{Y}$3, $\hat{Y}$4, to represent candidate map elements,
• point queries shared across instances to decode each element to $\hat{Y}$5 points,
• auxiliary foreground queries for segmentation.

Each of the six decoder layers has three sub-modules: self-attention among all queries, cross-attention between queries and encoder outputs, and intra-instance self-attention among point queries of the same instance. The decoder outputs are produced by a classification head for category scores $\hat{Y}$6, a point regression head for normalized coordinates $\hat{Y}$7, and a segmentation branch that predicts a bird’s-eye-view semantic map.

This architecture is explicitly designed to reason jointly about semantic and geometric consistency across trips. That design choice distinguishes it from pipelines that aggregate vehicle outputs after independent perception rather than within a learned fusion module.

3. Matching and optimization formulation

A central technical component of EGC-VMAP is hierarchical matching, used to assign predicted elements to ground truth during training. The process has two stages: instance-level matching and point-level matching.

Instance-level matching

Let $\hat{Y}$8 be a permutation over the $\hat{Y}$9 predictions. The framework minimizes

$E_k$0

with

$E_k$1

The use of Hungarian matching at the instance level enforces a one-to-one correspondence between predictions and ground-truth elements.

Point-level matching

For each matched pair $E_k$2, the method finds the best permutation $E_k$3 over its $E_k$4 points by minimizing

$E_k$5

This second stage provides the best pairing of vertices within each polyline or polygon.

Multi-objective loss

The total loss is defined as

$E_k$6

Here, $E_k$7 are hyper-parameters balancing the individual terms.

The classification loss is

$E_k$8

The point-to-point loss is the Manhattan distance over matched points,

$E_k$9

The edge direction loss encourages predicted edge vectors $k = 0 \ldots N_e - 1$0 to align with ground-truth edges $k = 0 \ldots N_e - 1$1 using cosine similarity,

$k = 0 \ldots N_e - 1$2

The segmentation term is standard pixel-wise cross-entropy on the auxiliary segmentation output (Feng et al., 11 Jul 2025).

The stated role of hierarchical matching is not only assignment accuracy but training stability: it greatly reduces spurious assignments, stabilizes gradient flow for both classification and geometry losses, and accelerates convergence. This suggests that EGC-VMAP treats correspondence uncertainty as a first-order optimization issue rather than a post hoc bookkeeping problem.

4. Data preprocessing, training, and inference

The training and inference pipeline is specified as a sequence of preprocessing, fusion, decoding, and postprocessing operations (Feng et al., 11 Jul 2025).

Preprocessing

The framework gathers $k = 0 \ldots N_e - 1$3 asynchronous, crowdsourced perceived maps over the same region. Each vector element is resampled by arc length to $k = 0 \ldots N_e - 1$4 points. Coordinates are min-max normalized into a fixed spatial window, with $k = 0 \ldots N_e - 1$5 given as an example.

Stable training via hierarchical matching

The two-stage matching procedure is used during supervised learning. Instance-level Hungarian matching establishes one-to-one assignment, and point-level matching then pairs vertices within each matched structure. The paper attributes stable gradient flow and accelerated convergence to this decomposition.

Inference and map fusion

Given new crowdsourced vectorized maps $k = 0 \ldots N_e - 1$6, the inference pipeline is:

1. apply the same preprocessing, including resampling and normalization;
2. encode with the Trip-Aware Transformer to produce query embeddings;
3. decode to obtain predicted elements $k = 0 \ldots N_e - 1$7;
4. filter low-confidence elements by thresholding $k = 0 \ldots N_e - 1$8;
5. denormalize coordinates to real-world WGS84 or local map coordinates;
6. aggregate overlapping predictions, for example by simple NMS or clustering, to emit the unified city-scale map.

This sequence reveals an important architectural boundary: the learned core ends at vector prediction, while the final city-scale consolidation can still use conventional aggregation mechanisms. A plausible implication is that the framework combines learned fusion with practical map-production postprocessing rather than replacing the latter entirely.

5. Evaluation protocol and empirical results

EGC-VMAP is validated on the Navinfo Dataset, described as open-sourced and containing 10,000 scenes across six cities: Beijing, Shanghai, Guangzhou, Shenzhen, Chongqing, and Tangshan. Each scene contains at least 10 distinct crowdsourced vehicles over multiple time points. The semantic element classes are lane dividers, stop lines, and crosswalks. Ground truth is obtained from specialized LiDAR survey and manual annotation (Feng et al., 11 Jul 2025).

Evaluation metrics

The paper evaluates geometric and detection quality using Chamfer Distance $k = 0 \ldots N_e - 1$9 between predicted point sets and ground-truth points, and AP at thresholds $N_v$0:

$N_v$1

Category-wise metrics are $N_v$2, $N_v$3, and $N_v$4, and mean average precision is

$N_v$5

Quantitative performance

Setting	mAP	Reported gain
StreamMapNet	40.71	—
StreamMapNet + EGC-VMAP	45.52	+4.81
Single-vehicle fusion (10 samples)	41.52	—
Crowdsourced fusion (10 vehicles)	45.52	+4.00
HDMapNet	18.28	—
HDMapNet + EGC-VMAP	23.20	+4.92
VectorMapNet	23.00	—
VectorMapNet + EGC-VMAP	28.63	+5.63
MapTR	34.35	—
MapTR + EGC-VMAP	38.43	+4.08

Under adverse conditions, the reported gains relative to StreamMapNet are +4.92 mAP in rain and +5.96 mAP at night. Qualitative examples are reported for challenging intersections, sharp turns, and occluded regions, where EGC-VMAP exhibits more complete, geometrically consistent polylines than single-vehicle methods.

The empirical pattern is consistent across both baseline replacement and fusion comparison. When compared with single-vehicle fusion using 10 samples, crowdsourced fusion using 10 vehicles improves mAP from 41.52 to 45.52. This suggests that the measured benefit is not reducible to observing more data from one platform; rather, the multi-vehicle, multi-temporal setting appears to contribute materially to the reported robustness.

6. Cost, scope, limitations, and future directions

A central claimed outcome is cost reduction. By replacing LiDAR survey and manual labeling with crowdsourced aggregation and end-to-end learning, EGC-VMAP reports an approximately 90% reduction in manual annotation cost at city scale (Feng et al., 11 Jul 2025). In the framing of the paper, this cost property is inseparable from the architectural decision to compute lightweight per-vehicle vectorized maps on the cars and aggregate them in the cloud.

The strengths identified for the framework are cost-effective, scalable city-wide HD mapping without specialized vehicles; robust geometric consistency through redundant multi-vehicle, multi-temporal fusion; and an end-to-end differentiable design that avoids brittle multi-stage pipelines. These strengths are directly aligned with the shortcomings attributed to offline LiDAR surveys and single-vehicle perception.

The limitations are also explicit. EGC-VMAP still relies on sufficient spatial and temporal coverage by crowdsourced vehicles, so rare roads or new constructions may be under-observed. Inference on very large areas may require region-by-region tiling and stitching because of latency. The semantic scope is currently limited to a fixed set of element types. These constraints are important when interpreting the framework’s city-scale claim: the system is scalable in the sense reported by the paper, but not coverage-independent.

The future directions listed include adaptive query allocation for dynamic numbers of elements per region, real-time incremental updates as new vehicle trips arrive, extension to 3D map elements such as curbs, barriers, and overpasses, and uncertainty estimation for active data collection in under-mapped regions. A plausible implication is that EGC-VMAP is best understood as a foundation for learned map fusion from vehicle fleets rather than as a completed universal mapping stack. Within that scope, its main contribution is to formalize city-scale vector map generation as a learned fusion problem over crowdsourced vector elements, with Trip-Aware Transformer fusion, hierarchical matching, and multi-task supervision as the key technical mechanisms.