MapDreamer: Aerial Imagery Conditioned Latent Diffusion for Lane-Level Map Generation

2026-07-01

AI Summary

This work addresses the scalability bottleneck in high-definition map generation caused by the heavy reliance on manual annotation for lane-level topological modeling. The authors propose a method that automatically produces explicit, vectorized topological maps from a single aerial image. Key innovations include a lane cardinality module and background ghost-lane latent variables to mitigate slot collapse, along with a sliding-window global graph aggregation strategy enabling seamless city-scale map stitching. The model integrates a variational autoencoder, Transformer architecture, latent diffusion mechanism, and cross-attention modules to effectively capture both geometric and topological relationships. Evaluated on the UrbanLaneGraph dataset, the approach significantly outperforms non-generative baselines in terms of geometric accuracy and topological fidelity.

Abstract

High definition map generation is essential for autonomous driving, yet remains a labor-intensive process at scale. We present MapDreamer, a generative diffusion model that synthesizes lane-level vector maps with explicit topology directly from a single aerial image. MapDreamer learns a compact latent representation of lane centerlines and their topological relations using a variational autoencoder and predicts graphs with a transformer-based latent diffusion model. To align generated maps with the observed scene, we condition each denoising step on dense aerial features injected through cross-attention. To handle the varying number of lanes across scenes, we propose a lane cardinality module paired with background ghost lane latents, a learned buffer that prevents slot collapse during diffusion. Furthermore, we introduce a sliding-window global graph aggregation strategy that stitches local tiles into city-scale maps while preserving connectivity through encoded lane boundaries. Experiments on UrbanLaneGraph derived from Argoverse 2 show improved geometric and topological fidelity over non-generative baselines.

Problem

• high definition map generation
• autonomous driving
• lane-level map
• aerial imagery
• vector map synthesis

Innovation

• latent diffusion
• lane-level map generation
• topological graph prediction
• cross-attention conditioning
• sliding-window aggregation