ReSMap: Recasting Satellite Priors for Robust and Accurate Online HD Map Construction
Authors: Kyungmin Kim, Sumin Lee, Sungoh Jeong, DoHyun Lim, Seonghyun Park, Soonmin Hwang
Organizations: Department of Automotive Engineering (Automotive-Computer Convergence), Hanyang University, Seoul, South Korea · Department of Computer Science, Hanyang University, Seoul, South Korea · Department of Electrical Engineering, Hanyang University, Seoul, South Korea · Department of Automotive Engineering, Hanyang University, Seoul, South Korea
Despite recent advances in camera-based online HD map construction, these systems remain susceptible to sensor failures due to their reliance on onboard cameras. In contrast, satellite images remain available independently of real-time sensing conditions as they are cached offline. Existing methods utilizing satellite images, however, primarily leverage them under nominal conditions, without explicitly addressing them as redundancy when a sensor degrades. To address this gap, we propose ReSMap, a redundancy-aware camera-satellite fusion framework for robust online HD map construction. Our framework explicitly leverages satellite images as a complementary source to improve robustness against camera failures. A Predict-then-Fuse (PTF) module fuses camera and satellite BEV features by using each branch's own prediction confidence as a per-pixel reliability weight, with no learned router. An Object-Centric Cooperative Multi-modal Query (OCMQ) decoder then extends this per-modality separation to the query level: each modality cross-attends only to its own BEV, and per-modality updates are merged through a learned per-token gate. Our model achieves state-of-the-art performance on nuScenes across all splits and range settings. More importantly, it also sets the new state-of-the-art under camera-failure scenarios, demonstrating that satellite images serve not only as a strong prior but as practical redundancy for robust online HD mapping.
Figures & tables
Figure 1: Overview of the model architecture of ReSMap . Two modality-specific encoders produce camera and satellite BEV features, each with a BEV segmentation head that outputs per-pixel confidence. The two BEV streams are fused by a parameter-free, confidence-weighted gate (PTF). A single object-query set then gathers modality-specific updates from the camera, satellite, and fused BEV streams, which are recomposed by a learned per-token gate (OCMQ) and decoded into the final vector HD map.
Figure 2: Per-pixel and per-query gate response under camera dropout. (a) Camera and satellite reliability normalized as in Eq. ( 3 ), rˉm=rm/(rcam+rsat) , over active cells A={x:rcamclean(x)≥0.35} . Removing all cameras collapses rˉcam while rˉsat rises to cover the same regions. (b) The same routing is visible at the query level: sorting queries by αcamclean (top) and plotting the matched change Δαcam under all-camera drop (bottom) shows a negative shift for 98% of queries, with the largest drops on those most camera-reliant in the clean run.
Figure 3: Effect of query design on Hungarian matching. (a) With independent cam/sat/fuse query pools, bipartite matching routes each GT to the strongest modality only, causing class–modality entanglement (e.g. dividers and boundaries learned only from camera). (b) OCMQ uses a single object-query pool with modality-specific updates merged via a per-token gate, so each query absorbs cam, sat, and fuse gradients jointly and every class receives signal from all modalities.
Range
Method
Venue
Mod.
Temp.
Ep.
AP p
AP d
AP b
mAP
FPS
60 m × 30 m
MapTracker † [ 4 ]
ECCV’24
C
✓
58
45.9
30.0
45.1
40.3
13.5
MapTRv2 [ 5 ]
IJCV’24
C+L
×
110
18.8
29.5
61.7
36.6
7.3
DAMap [ 6 ]
ICCV’25
C+L
×
110
19.2
32.8
66.1
39.4
7.2
SDTagNet [ 9 ]
NeurIPS’25
C+SD
×
110
11.7
25.3
40.5
25.8
9.6
PriorDrive ∗ [ 10 ]
AAAI’26
C+SD+HD
✓
110
44.7
35.5
47.3
42.5
8.2
SatforHDMap [ 11 ]
ICRA’24
C+SA
×
110
44.1
12.1
24.8
27.0
14.1
Table 1: Online HD map construction performance on the geographically non-overlapping split of nuScenes [ 3 ] ; FPS measured on a single RTX 4090. Mod. : C, L, SD, HD, SA denote Camera, LiDAR, SD map, prior HD map, and Satellite images; Temp. : temporal memory or historical-prediction priors. ∗ from the original paper; † Consistent-GT supervision [ 4 ] .
Camera Drop
Method
Mod.
Temp.
Ep.
Clean
Front-1
Front-3
All
MapTracker † [ 4 ]
C
✓
58
40.3
23.0 (-42.9%)
10.8 (-73.2%)
0.1 (-99.8%)
MapTRv2 [ 5 ]
C+L
×
110
36.6
32.2 (-12.0%)
29.5 (-19.4%)
23.8 (-35.0%)
DAMap [ 6 ]
C+L
×
110
39.4
35.8 (-9.1%)
34.2 (-13.2%)
29.2 (-25.9%)
SDTagNet [ 9 ]
C+SD
×
110
25.8
19.8 (-23.3%)
14.5 (-43.8%)
4.0 (-84.5%)
SatforHDMap [ 11 ]
C+SA
×
110
27.0
21.5 (-20.4%)
21.3 (-21.1%)
14.3 (-47.0%)
Table 2: Robustness on the geographically non-overlapping split of nuScenes [ 3 ] , at 60m×30m . Front-1 , Front-3 , and All zero out the front camera, the three front cameras, and all six cameras at inference, respectively. Each cell shows mAP and the relative change from Clean, (Drop−Clean)/Clean×100% , in red. Symbols as in Tab. 1 .
Figure 4: Qualitative comparison of four methods. Top: multi-view cameras and aerial satellite images for the two evaluation ROIs ( 60m×30m and 100m×50m ). Bottom: predicted vector maps under Clean , Front-3 at 60m×30m , and Clean at 100m×50m (forward axis horizontal, ego facing right).
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Venue
Mod.
Temp.
Ep.
AP p
AP d
AP b
mAP
InteractionMap ∗ [ 48 ]
CVPR’25
C
✓
110
75.2
75.0
76.1
75.5
MapTracker † [ 4 ]
ECCV’24
C
✓
70
80.0
74.1
74.1
76.1
MapTRv2 [ 5 ]
IJCV’24
C+L
×
24
65.6
66.5
74.8
69.0
Mask2Map ∗ [ 7 ]
ECCV’24
C+L
×
24
76.5
76.6
82.1
78.4
DAMap [ 6 ]
ICCV’25
C+L
×
24
67.7
72.4
77.1
72.4
SDTagNet [ 9 ]
NeurIPS’25
C+SD
×
30
61.5
63.1
61.4
62.0
Appendix
Table 4: Online HD map construction performance on the original split of nuScenes, 60m×30m . Mod. : C, L, SD, HD, SA denote Camera, LiDAR, SD map, prior HD map, and Satellite images; Temp. : temporal memory or historical-prediction priors. ∗ from the original paper; † Consistent-GT supervision [ 4 ] . PriorDrive uses an SD map together with a noisy prior HD map and online local maps from earlier traversals.
Camera Drop
Method
Mod.
Temp.
AP p
AP d
AP b
Clean
Front-1
Front-3
All
MapTracker † [ 4 ]
C
✓
70.0
75.1
69.0
71.3
64.4 (-9.7%)
30.0 (-57.9%)
0.2 (-99.7%)
SDTagNet [ 9 ]
C+SD
×
64.3
70.8
68.4
67.8
64.6 (-4.7%)
52.3 (-22.9%)
18.6 (-72.6%)
SatforHDMap [ 11 ]
C+SA
×
17.6
40.5
37.3
31.8
30.0 (-5.7%)
29.3 (-7.9%)
21.0 (-34.0%)
ReSMap (Ours)
C+SA
×
69.2
67.3
65.6
67.4
63.5 (-5.8%)
57.1 (-15.3%)
46.9 (-30.4%)
ReSMap (Ours)
C+SA
✓
77.3
75.8
74.3
75.8
71.2 (-6.1%)
58.5 (-22.8%)
30.9 (-59.2%)
Appendix
Table 5: Results on the geographically non-overlapping split of Argoverse 2 [ 45 ] , at 60m×30m , with satellite tiles from Google Maps. Left: clean inputs. Right: camera dropout, where Front-1 , Front-3 , and All zero out the front-center camera, the three front cameras, and all seven ring cameras at inference. Each drop cell shows mAP and the relative change from Clean in red. † Consistent-GT supervision [ 4 ] .
Camera Drop
Corruption
Method
Mod.
Temp.
Ep.
Clean
Front-1
Front-3
All
mRR ↑
mCE ↓
MapTracker † [ 4 ]
C
✓
70
76.1
41.1 (-46.0%)
17.0 (-77.7%)
0.1 (-99.9%)
69.1
61.2
SafeMap ∗ [ 29 ]
C
×
24
52.5
42.4 (-19.2%)
–
–
51.2
90.6
MapTRv2 [ 5 ]
C+L
×
24
69.0
62.3 (-9.7%)
55.5 (-19.6%)
41.3 (-40.1%)
80.7
55.4
DAMap [ 6 ]
C+L
×
24
72.4
62.4 (-13.8%)
54.8 (-24.3%)
35.3 (-51.2%)
79.3
55.8
SDTagNet [ 9 ]
C+SD
×
30
62.0
47.1 (-24.0%)
38.7 (-37.6%)
14.5 (-76.6%)
68.6
75.6
Appendix
Table 6: Robustness on the original split of nuScenes, at 60m×30m . Front-1 , Front-3 , and All-camera zero out the front camera, the three front cameras, and all six cameras at inference, respectively. We also report MapBench [ 47 ] mRR and mCE (in %), with MapTR [ 1 ] as the mCE reference following MapBench. Each camera-drop cell shows mAP and the relative change from Clean, (Drop−Clean)/Clean×100% , in red. SafeMap [ 29 ] only reports single-view drop. Temp. : temporal memory or historical-prediction priors. ∗ from the original paper; † Consistent-GT supervision [ 4 ] .
Method
Mod.
mRR ↑
mCE ↓
MapTracker † [ 4 ]
C
65.7
1.000
MapTRv2 [ 5 ]
C+L
87.3
0.927
DAMap [ 6 ]
C+L
85.5
0.920
SDTagNet [ 9 ]
C+SD
66.9
1.132
SatforHDMap [ 11 ]
C+SA
90.4
1.052
ReSMap (Ours)
C+SA
90.8
0.712
Appendix
Table 7: MapBench [ 47 ] sensor-corruption robustness on the geographically non-overlapping split of nuScenes [ 3 ] , at 60m×30m : mean Resilience Rate (mRR, clean-normalized retention over the 24 corruptions) and mean Corruption Error (mCE, relative to MapTracker). ReSMap is the temporal model. † Consistent-GT supervision [ 4 ] .
Method
BEV-feature dropout
Clean
Front-1
Front-3
All
SatforHDMap [ 11 ]
no
27.0
21.5
21.3
14.3
SatforHDMap [ 11 ]
yes
27.6
24.9
21.0
16.7
SDTagNet [ 9 ]
no
25.8
19.8
14.5
4.0
SDTagNet [ 9 ]
yes
26.3
20.1
13.0
4.3
DAMap [ 6 ]
no
39.4
35.8
34.2
29.2
DAMap [ 6 ]
yes
38.3
35.2
34.5
30.7
Appendix
Table 8: BEV-feature dropout on and off. † Our model before hyperparameter tuning.
Satellite Corruption mAP
Method
Mod.
Temp.
Ep.
Clean
Temporal swap
Geometric noise
ReSMap (Ours)
C+SA
×
24
47.3
43.1 (-8.9%)
43.4 (-8.2%)
ReSMap (Ours)
C+SA
✓
24
54.9
51.6 (-6.0%)
51.9 (-5.5%)
Appendix
Table 9: Robustness to satellite image corruption on the geographically non-overlapping split of nuScenes [ 3 ] , at 60m×30m , with corruption probability p=0.2 . Temporal swap: feeding the t±1 tile in place of the t -frame tile. Geometric noise: shifting the tile by a Gaussian offset ( σt=1m ) and/or rotating it ( σr=5∘ ); for each corrupted sample we randomly pick one of {shift, rotate, both}. Each cell shows mAP and the relative change from Clean, (Corrupt−Clean)/Clean×100% , in red.
Setting
Branch / Input
AP p
AP d
AP b
mAP
(a) Per-branch prediction on clean inputs
Camera
11.0
17.2
38.6
22.3
Satellite
17.8
2.7
9.7
10.1
Fused
58.0
11.5
4.6
24.7
(b) Final prediction under camera dropout
Clean
67.8
28.1
51.0
49.0
Front-3 drop
58.0 (-14.5%)
19.2 (-31.7%)
22.1 (-56.7%)
33.1 (-32.4%)
Appendix
Table 10: Per-branch prediction on clean inputs and final prediction under Front-3 drop (three front cameras zeroed).
Setting
rcam
rsat
rfuse
∥Ffuse∥
Temporal
Clean
0.751
0.473
0.678
11.59
Front-1 drop
0.497 (−0.254)
0.473
0.575 (−0.103)
10.67 (−0.92)
All-cam drop
0.166 (−0.585)
0.473
0.437 (−0.242)
9.37 (−2.21)
Non-temporal
Clean
0.686
0.456
0.545
9.78
Front-1 drop
0.553 (−0.133)
0.456
0.507 (−0.038)
9.56 (−0.22)
All-cam drop
0.278 (−0.408)
0.456
0.445 (−0.100)
9.17 (−0.61)
Appendix
Table 11: Diagnosing temporal vs. non-temporal robustness on the geographically non-overlapping split validation set (5981 frames). All BEV-pixel statistics are restricted to the active cells At={x:rcamclean(x,t)≥0.35} . rcam,rsat follow Eq. ( 2 ); rfuse is an auxiliary probe ( Segfuse on Ffuse ); ∥Ffuse∥ is the channel-wise L2 norm of Ffuse . Parentheses are absolute changes from the clean row.
Figure 5: Additional qualitative comparison on boston-seaport.
Figure 6: Additional qualitative comparison on singapore-hollandvillage.
Figure 7: Additional qualitative comparison on boston-seaport.
Figure 8: Additional qualitative comparison on singapore-queenstown.
Figure 9: Additional qualitative comparison on singapore-onenorth. The extended 100×50m ROI setting illustrates a representative failure case in the lower-left region, where occlusion and visual ambiguity in the satellite image caused by nearby buildings may lead to missed map elements.
High-definition (HD) maps provide essential semantic information about road structures for autonomous driving, but existing HD map construction methods typically require calibrated multi-camera rigs and explicit 2D-to-BEV transformations. Such pipelines degrade when camera views are missing or pose estimates are inaccurate, limiting their use across heterogeneous fleet configurations. We introduce FlexMap, a vectorized HD mapping framework that adapts to varying camera configurations without architectural changes or per-configuration retraining and does not require camera parameters as model input. FlexMap replaces explicit geometric projection with a geometry foundation model that encodes cross-view 3D structure. A spatial-temporal enhancement module then separates cross-view spatial reasoning from temporal aggregation, while a camera-aware decoder uses the token produced for each input view to adapt its attention without camera poses. Experiments on nuScenes and Argoverse 2 show that FlexMap outperforms pose-dependent baselines using estimated poses and maintains comparable accuracy across all evaluated camera configurations, including those with missing views.
Run Wang, Chaoyi Zhou, Amir Salarpour +5
School of Computing, Clemson University · 2Tacoma School of Engineering and Technology, University of Washington
High-resolution aerial imagery has recently emerged as a complementary modality for automated driving perception and has shown potential to improve birds-eye-view (BEV) scene understanding when fused with onboard sensors. Prior work demonstrated performance gains for online high-definition (HD) map construction through aerial-onboard fusion; however, conventional end-to-end fusion does not fully exploit the structural information contained in aerial representations. In this work, we introduce AerialFusionMapNet, a fusion-based mapping framework with a structured two-stage training strategy that explicitly enhances the contribution of aerial features within a unified pipeline. The proposed training scheme enables more effective integration of structural aerial priors. On the nuScenes geographic split, AerialFusionMapNet achieves up to 54.7 mAP, improving over prior aerial-onboard fusion baselines from 48.8 mAP by +5.9 absolute and +12.1% relative. The results suggest that structured training design, rather than increased architectural complexity, plays a more decisive role in unlocking the full potential of aerial imagery for online HD map construction. Code and trained models are available at https://github.com/DriverlessMobility/AerialFusionMapNet.
Daniel Lengerer, Mathias Pechinger, Klaus Bogenberger +1
Technical University of Applied Sciences Augsburg, Germany. · Technical University of Munich, Germany.
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.