SyncBEV: Diffusion-Based, High-Fidelity Synchronized Bird's-Eye-View Synthesis from Ground-View Data

Kurzfassung

Bird's Eye View (BEV) representations make it much easier to understand driving scenes, allowing for straightforward measurement of distances to vehicles, road boundaries, and free space from a top-down perspective. This simplification supports safer navigation and more effective planning in autonomous vehicles. However, generating high-quality BEV images from ground-level sensors remains challenging. Most cross-view synthesis methods struggle with generation fidelity and face a fundamental limitation in their training data: aerial and ground images are often captured at different times, meaning that dynamic objects like vehicles or pedestrians often appear in one view but not the other. This leads to mismatches in the ground truth labels that models are trained and evaluated against. In this work, we introduce SyncBEV, a BEV synthesis method built on Stable Diffusion XL, a powerful pretrained diffusion model. Our approach uses structural projections from monocular depth and incorporates rotation augmentation for robust viewpoint handling. The result is BEV imagery that is not only more realistic and detailed than previous baselines like GOLD-BEV, but also better aligned with the real-time scene. Notably, when tested without any fine-tuning on the KITTI-CVL dataset, our model generates BEV images that capture dynamic objects missed by the dataset’s own asynchronous aerial ground truth to a fair extent. This demonstrates that our method can produce top-down views that are more temporally faithful to the actual scene than traditional aerial references.