High-definition (HD) maps are crucial to autonomous driving, providing structured representations of road elements to support navigation and planning. However, existing query-based methods often employ random query initialization and depend on implicit temporal modeling, which lead to temporal inconsistencies and instabilities during the construction of a global map. To overcome these challenges, we introduce a novel end-to-end framework for consistent online HD vectorized map construction, which jointly performs map instance tracking and short-term prediction. First, we propose a Semantic-Aware Query Generator that initializes queries with spatially aligned semantic masks to capture scene-level context globally. Next, we design a History Rasterized Map Memory to store fine-grained instance-level maps for each tracked instance, enabling explicit historical priors. A History-Map Guidance Module then integrates rasterized map information into track queries, improving temporal continuity. Finally, we propose a Short-Term Future Guidance module to forecast the immediate motion of map instances based on the stored history trajectories. These predicted future locations serve as hints for tracked instances to further avoid implausible predictions and keep temporal consistency. Extensive experiments on the nuScenes and Argoverse2 datasets demonstrate that our proposed method outperforms state-of-the-art (SOTA) methods with good efficiency.
Figures & tables
Figure 1: Comparison of several HD map construction methods: (a) Single-Frame, (b) Temporal Propagation, (c) PredMapNet (Ours) utilizes semantic-aware information and enables both historical & future reasoning for better query decoding.
Figure 2: The architecture pipeline of PredMapNet. At each frame, multi-view images are processed by a BEV encoder to extract perception features. The Semantic-Aware Query Generator 3.2 produces semantic-aware detection queries and rasterized map from BEV features. A History Rasterized Map Memory 3.3 is maintained to store instance-level segmentation masks over time. The History-Map Guidance Module 3.4 refines track queries with historical geometric priors from memory. Simultaneously, the Short-Term Future Guidance Module 3.5 predicts future polylines from historical trajectories and fuses them into track queries to guide query initialization in the next frame. Together, these modules enable temporally consistent and robust map instance construction across frames.
Method
Epoch
APped
APdivider
APboundary
m AP ↑
C- m AP ↑
FPS
PivotNet [ 9 ]
24
56.2
56.5
60.1
57.6
-
-
MapTRv2 [ 19 ]
24
59.8
62.4
62.4
61.5
41.7
14.1
HRMapNet [ 43 ]
24
65.8
67.4
68.5
67.3
49.2
10.3
StreamMapNet [ 39 ]
24
61.9
66.3
62.1
63.4
38.4
13.1
MGMap [ 22 ]
24
61.8
65.0
67.5
64.8
43.5
13.4
Mask2Map [ 8 ]
24
70.6
71.3
72.9
71.6
55.8
9.2
Table 1: Comparison with SOTA methods on the nuScenes validation set (old split [ 19 ] ).
Method
Epoch
APped
APdivider
APboundary
m AP ↑
C- m AP ↑
MapTRv2 [ 19 ]
24
62.9
72.1
67.1
67.4
-
HRMapNet [ 43 ]
30
65.1
71.4
68.6
68.3
-
StreamMapNet [ 39 ]
72
70.5
74.2
66.1
70.3
-
Mask2Map [ 8 ]
24
68.1
72.7
73.7
71.5
-
MGMapNet [ 38 ]
24
71.3
76.0
73.1
73.6
-
MapTracker [ 5 ]
35
76.9
79.9
73.6
76.8
68.3
Table 2: Comparison with SOTA methods on the Argoverse2 validation set.
Method
Epoch
APped†
APdivider†
APboundary†
mAP†
MapTR [ 18 ]
24
32.4
23.5
17.1
24.3
MapTRv2⋆ [ 19 ]
24
49.9
34.7
25.7
36.7
Mask2Map [ 8 ]
24
62.9
52.3
48.9
54.7
PredMapNet (Ours)
24
69.3
60.1
63.5
64.3
Table 3: Comparison of SOTA methods on nuScenes validation set (old split [ 19 ] ) with rasterization-based metric. The ” ⋆ ” indicates results reproduced using public codes.
Figure 3: Qualitative visualization on nuScenes val set.
Dataset
Method
m AP ↑
C- m AP ↑
nuScenes
StreamMapNet [ 39 ]
33.5
22.2
MapTracker [ 5 ]
40.3
32.5
PredMapNet (Ours)
42.1
33.7
Argoverse2
StreamMapNet [ 39 ]
64.4
54.4
MapTracker [ 5 ]
70.3
61.3
PredMapNet (Ours)
71.2
62.3
Table 4: Comparisons on non-overlapping datasets.
APped
APdivider
APboundary
m AP ↑
Baseline
77.3
72.4
74.2
74.7
+Semantic-Aware Query 3.2
77.2
73.1
74.7
75.0
+History-Map Guidance 3.4
77.6
73.6
75.0
75.4
+Short-Term Future Guidance 3.5
78.1
74.7
76.1
76.3
+Aux Depth Supervision
79.0
75.2
76.5
76.9
Table 5: Ablation study of main components of PredMapNet
Constructing reliable online HD maps remains challenging in dynamic urban environments due to moving objects and occlusions. While recent works employ feature-level temporal fusion to address this, they rely solely on per-frame ground truth supervision. Consequently, they lack an explicit objective to directly penalize the geometric noise and temporal jitter between consecutive online HD maps. To address this, we propose MapTCL, an auxiliary training strategy that formulates temporal consistency loss between current and past frames via bidirectional alignment. Specifically, Bidirectional Vector Consistency Learning (BVCL) models the geometric and semantic discrepancies between associated past and current vector instances as an auxiliary loss. We also employ Raster map Consistency Learning (RCL) as an additional loss to stabilize dense BEV features. By jointly training with these dual losses, MapTCL improves the temporal stability of generated HD maps. Extensive experiments on two standard benchmarks demonstrate the effectiveness of our approach. As a versatile plug-and-play module, MapTCL consistently enhances existing baseline models, achieving gains of +3.7 mAP & +2.8 C-mAP on nuScenes and +3.1 mAP & +2.5 C-mAP on Argoverse 2 without additional inference overhead.
Hyeonseo Kim, Juyeb Shin, Hyeonjun Jeong +2
Robotics Program, Korea Advanced Institute of Science and Technology (KAIST), Daejeon 34141, South Korea · Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology (KAIST), Daejeon 34141, South Korea
High-definition (HD) maps are essential for autonomous driving systems. In constructing such maps, onboard multi-view camera images, standard-definition maps and satellite images provide crucial information. However, due to the modality and perspective differences among these data sources, existing methods often struggle to effectively align and fuse them, making online HD map construction still challenging. To address these issues, we propose Driver2Map, an online HD map construction model inspired by human drivers. Unlike existing HD map construction models that utilize only two modalities, our Driver2Map can simultaneously exploit three modalities. Specifically, we propose a "two-stage alignment" strategy to reduce spatial misalignment across different modalities. Additionally, we introduce "Pose-Guided BEV Fusion", a BEV (bird's-eye-view) generation module that leverages camera pose information to adaptively weight multi-view features, thereby effectively suppressing cross-view feature overlap during BEV generation. Also, we design a "Pretrained Prior for Map Refinement" module to refine the initial prediction by learning map structure priors, thus improving the HD map prediction under dynamic occlusions. Extensive experiments demonstrate that Driver2Map outperforms existing methods on both IoU and AP metrics.
Accurate High-Definition (HD) map construction is critical for autonomous driving, yet existing methods face a fundamental trade-off: vectorization-based approaches preserve topology but struggle with geometric fidelity, while rasterization-based approaches enable precise geometric supervision but produce unstructured outputs. To bridge this gap, we propose GSMap, a novel framework that unifies both paradigms via a learnable 2D Gaussian representation. Each map element is modeled as an ordered sequence of 2D Gaussians, whose centers correspond to the vertices of the vectorized polyline/polygon. This formulation enables simultaneous optimization through: (1) Differentiable rasterization that enforces pixel-level geometric constraints, and (2) Topology-aware vectorization that maintains structural regularity. Experiments on both nuScenes and Argoverse2 demonstrate that our Gaussian-based representation effectively unifies geometric and topological learning, achieving significant performance improvements and demonstrating strong compatibility with existing HD mapping architectures. Code will be available at https://github.com/peakpang/GSMap
Zhenxuan Zeng, Lingxuan Wang, Sheng Yang +4
School of Computer Science, Northwestern Polytechnical University, China · Unmanned Vehicle Dept, Cainiao Inc., Alibaba Group, China