The DiffVL framework redefines visual localization as a diffusion‑based GPS denoising task, using BEV‑conditioned visual cues and standard SD maps to achieve sub‑meter accuracy without high‑definition maps, and demonstrates its superiority through extensive autonomous‑driving experiments.
By fusing LiDAR point clouds and camera images into a unified bird‑eye‑view space and applying Transformer‑based perception, multi‑sensor fusion, and graph‑diffusion modeling, the proposed BEV+Transformer framework automatically detects and smooths ground‑level line features and signs for high‑definition maps with centimeter‑level accuracy, boosting production efficiency and reducing cost.
Vision‑based Bird’s‑Eye‑View (BEV) transforms camera data into a scale‑invariant, geometry‑friendly top‑down map using perspective‑transformation modules such as Lift‑Splat‑Shoot and Pseudo‑LiDAR, incorporates deformable convolutions and attention, and underpins modern autonomous‑driving detectors like BEVDet, BEVDepth, Detr3D, BEVFormer and PETR, while future research targets depth‑estimation bottlenecks, multimodal transformer fusion, and foundation‑model generalization.
