Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SatMap: Revisiting Satellite Maps as Prior for Online HD Map Construction

Kanak Mazumder, Fabian B. Flohr

TL;DR

SatMap tackles online HD map construction by integrating a global satellite-map prior into a multi-view camera BEV framework to produce vectorized map predictions. It employs a Swin Transformer-based satellite feature extractor, GFPN fusion, and a ConvFuser to merge satellite and camera BEV features, followed by a DETR-based decoder to output polylines for lanes, boundaries, and crossings. On nuScenes, SatMap delivers substantial gains over camera-only and camera-LiDAR baselines and demonstrates robustness in long-range and adverse weather, by mitigating depth ambiguity and occlusion through satellite context. The work highlights the practical value of satellite priors for robust, real-time HD map construction and points to future directions in temporal fusion and monocular ego-centric mapping.

Abstract

Online high-definition (HD) map construction is an essential part of a safe and robust end-to-end autonomous driving (AD) pipeline. Onboard camera-based approaches suffer from limited depth perception and degraded accuracy due to occlusion. In this work, we propose SatMap, an online vectorized HD map estimation method that integrates satellite maps with multi-view camera observations and directly predicts a vectorized HD map for downstream prediction and planning modules. Our method leverages lane-level semantics and texture from satellite imagery captured from a Bird's Eye View (BEV) perspective as a global prior, effectively mitigating depth ambiguity and occlusion. In our experiments on the nuScenes dataset, SatMap achieves 34.8% mAP performance improvement over the camera-only baseline and 8.5% mAP improvement over the camera-LiDAR fusion baseline. Moreover, we evaluate our model in long-range and adverse weather conditions to demonstrate the advantages of using a satellite prior map. Source code will be available at https://iv.ee.hm.edu/satmap/.

SatMap: Revisiting Satellite Maps as Prior for Online HD Map Construction

TL;DR

SatMap tackles online HD map construction by integrating a global satellite-map prior into a multi-view camera BEV framework to produce vectorized map predictions. It employs a Swin Transformer-based satellite feature extractor, GFPN fusion, and a ConvFuser to merge satellite and camera BEV features, followed by a DETR-based decoder to output polylines for lanes, boundaries, and crossings. On nuScenes, SatMap delivers substantial gains over camera-only and camera-LiDAR baselines and demonstrates robustness in long-range and adverse weather, by mitigating depth ambiguity and occlusion through satellite context. The work highlights the practical value of satellite priors for robust, real-time HD map construction and points to future directions in temporal fusion and monocular ego-centric mapping.

Abstract

Online high-definition (HD) map construction is an essential part of a safe and robust end-to-end autonomous driving (AD) pipeline. Onboard camera-based approaches suffer from limited depth perception and degraded accuracy due to occlusion. In this work, we propose SatMap, an online vectorized HD map estimation method that integrates satellite maps with multi-view camera observations and directly predicts a vectorized HD map for downstream prediction and planning modules. Our method leverages lane-level semantics and texture from satellite imagery captured from a Bird's Eye View (BEV) perspective as a global prior, effectively mitigating depth ambiguity and occlusion. In our experiments on the nuScenes dataset, SatMap achieves 34.8% mAP performance improvement over the camera-only baseline and 8.5% mAP improvement over the camera-LiDAR fusion baseline. Moreover, we evaluate our model in long-range and adverse weather conditions to demonstrate the advantages of using a satellite prior map. Source code will be available at https://iv.ee.hm.edu/satmap/.
Paper Structure (20 sections, 3 equations, 3 figures, 3 tables)

This paper contains 20 sections, 3 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Online Vectorized Map Learning
  4. Cross-view BEV learning
  5. Prior-based Map Learning
  6. Methodology
  7. Overview
  8. Multi-modal Feature Extraction
  9. Image Feature Extraction
  10. Satellite Map Fetching
  11. Satellite Feature Extraction
  12. BEV Feature Fusion
  13. Map Decoder
  14. Experiments
  15. Dataset
  16. ...and 5 more sections

Figures (3)

  • Figure 1: We propose SatMap, a camera-satellite fusion-based online vectorized HD map prediction framework. Due to occlusion in perspective view, camera-only models fail to predict accurate map polylines. With a satellite map as prior, SatMap can augment the occluded region with prior information and predict a much accurate HD map. The occluded areas are highlighted in red and magenta.
  • Figure 2: The architecture of our proposed SatMap architecture. First, multi-view image features are extracted and projected to BEV space, which are fused with BEV features extracted from the satellite map. The fused BEV feature is used for decoding vectorized map elements.
  • Figure 3: Visualization of qualitative results of SatMap in challenging scenes from the nuScenes validation dataset. The left column is the surround views, the middle column is the satellite map, the next columns are inference results of baseline and SatMap, and the right column is the corresponding ground truth. Green lines indicate boundaries, yellow lines indicate lane dividers, and blue lines indicate pedestrian crossings. The challenging regions are indicated in red.