Uni-PrevPredMap: Extending PrevPredMap to a Unified Framework of Prior-Informed Modeling for Online Vectorized HD Map Construction

Safety-critical autonomous driving motivates the effective use of prior information. For online vectorized HD map construction, temporal predictions and cost-efficient HD map priors are two complementary yet individually imperfect sources. However, existing prior-informed approaches typically use only one of them or assume the HD map prior to be reliable. We present Uni-PrevPredMap, a unified framework that treats both as imperfect priors and processes them within a single pipeline through their common vectorized representation. At its core is a tri-mode training paradigm that exposes the model to non-prior, temporal-prior, and temporal-map-fusion conditions. This single design enables one model to perceive reliably without any prior, exploit priors when available, and stay robust when priors are imperfect, rather than being trained under a single fixed prior condition. Uni-PrevPredMap achieves state-of-the-art map-absent performance on nuScenes and Argoverse2. Combining both priors yields gains beyond either source alone, while robustness to imperfect priors is evaluated under synthetic perturbation types unseen during training, indicating that the model can benefit from map priors without over-relying on them. Code is available at https://github.com/pnnnnnnn/Uni-PrevPredMap.