DeepDualMapper: A Gated Fusion Network for Automatic Map Extraction Using Aerial Images and Trajectories

Automatic map generation is crucial for location‑based services, yet existing methods rely on a single data source—either remote sensing images or vehicle trajectories—leading to information loss. By fusing both modalities, map quality can be significantly improved.

The authors introduce DeepDualMapper , a gated fusion network presented at AAAI‑20, which merges aerial images and trajectory data to produce maps automatically. The model treats map generation as a pixel‑wise binary classification problem, extending the U‑Net architecture with a Gated Fusion Module (GFM) and a Densely‑Supervised Refinement (DSR) decoder.

The GFM first aligns features from the two sources using a 1×1 convolution, then linearly combines them with learned gate values that sum to one, allowing the network to select the more reliable source for each region.

The refinement module, inspired by residual learning, further enhances the fused output to produce smoother and more continuous road networks.

Experiments on datasets from Porto, Shanghai, and Singapore compare DeepDualMapper with trajectory‑only (TCI, KDE, COBWEB), image‑only (DeconvNet), and early‑fusion baselines. Using IoU and F1‑score, DeepDualMapper consistently achieves the highest performance across all cities.

Visualization of the gate values shows the module automatically prefers the data source with higher confidence, while the refinement module produces smoother, more accurate road maps.

In conclusion, DeepDualMapper effectively integrates aerial imagery and trajectory data through gated fusion and refinement, delivering superior map extraction performance demonstrated on multiple urban datasets.