PrevPredMap: Exploring Temporal Modeling with Previous Predictions for Online Vectorized HD Map Construction

Nan Peng, Xun Zhou, Mingming Wang, Xiaojun Yang, Songming Chen, Guisong ChenWACV 2025

TL;DR

This paper introduces PrevPredMap, a novel framework for constructing vectorized High-Definition (HD) maps in real-time by leveraging previous predictions as temporal information. By meticulously designing modules to encode and utilize these past predictions, PrevPredMap achieves state-of-the-art performance on challenging benchmarks, offering a more robust and efficient approach to dynamic map creation.

What Problem Does This Paper Address?

Creating and maintaining High-Definition (HD) maps for autonomous driving is a critical but labor-intensive task. Traditional methods are costly and struggle to keep maps up-to-date. Online HD map construction, which uses onboard sensors to perceive map elements in real-time, offers a more cost-effective solution. While recent advancements have moved from semantic segmentation to directly predicting vectorized map elements, effectively incorporating temporal information to handle occlusions and improve map quality remains a challenge. Existing temporal modeling approaches often rely on intermediate representations like BEV features or query features. This paper explores a less-traveled path: can previous predictions themselves serve as a powerful temporal signal for constructing current HD maps, offering a higher level of abstraction and potentially better integration with map priors?

Methods

PrevPredMap introduces a novel temporal modeling framework built upon three primary components:

  • BEV Feature Extractor: A standard component that processes multi-view images to extract Bird's-Eye-View (BEV) features.
  • Previous-Predictions-Based Query Generator: This is a core innovation. It takes previous map predictions (category, confidence, and location) and encodes them into distinct "content queries" (for category/confidence) and "position queries" (for location). This module employs a crucial dual-mode strategy: during training, it randomly alternates between using only current frame information (single-frame mode) and incorporating previous predictions (temporal mode). This ensures robust performance in both scenarios during inference.
  • Dynamic-Position-Query Decoder: This module refines the temporal modeling. Unlike static approaches where position queries remain fixed, this decoder dynamically updates position queries at each layer based on the predicted locations from the preceding layer. This ensures that the decoder always has access to the most up-to-date location information derived from temporal cues.

The framework also includes an enhanced single-frame baseline that uses a group-wise one-to-many branch to reduce memory overhead, making it more efficient.

Datasets

The proposed PrevPredMap was evaluated on two prominent large-scale datasets:

  • nuScenes: Provides 2D vectorized maps and 2Hz RGB images from multiple cameras.
  • Argoverse2: Offers 3D vectorized maps and 20Hz RGB images. For fair comparison with existing methods, the authors adjusted its sampling frequency to 2Hz and 10Hz.

Key Insights and Ideas

  • Predictions as Temporal Features: The paper convincingly demonstrates that previous predictions, not just intermediate features, can be a highly effective temporal signal for HD map construction. This offers a more abstract and potentially more robust representation.
  • Dual-Mode Training is Key: The random switching between single-frame and temporal modes during training is critical. It allows the model to maintain strong performance even when temporal information is unavailable (e.g., on the first frame), preventing a significant drop in accuracy.
  • Dynamic Position Queries Matter: The dynamic updating of position queries within the decoder is essential for effectively leveraging the temporal information encoded by the previous-predictions-based query generator. Without this, the location information could become stale.
  • Handling Occlusions: Qualitative results show PrevPredMap's superior ability to reconstruct map elements in challenging occlusion scenarios compared to single-frame methods, highlighting the practical benefit of temporal modeling.

Key Results

PrevPredMap achieves state-of-the-art results on both nuScenes and Argoverse2 datasets:

  • nuScenes: Achieved 66.3 mAP (24-epoch training) and 71.3 mAP (110-epoch training), outperforming existing methods in convergence, validation accuracy, and inference speed.
  • Argoverse2: Surpassed all SOTA methods in both 2D and 3D vectorized map construction, demonstrating its effectiveness on a dataset with richer information.
  • Component Ablation: The combination of the previous-predictions-based query generator and the dynamic-position-query decoder led to a significant 4.8 mAP improvement.
  • Single-Frame vs. Temporal: The single-frame mode of PrevPredMap is competitive with existing SOTA methods, and the temporal mode offers a notable improvement, with only a small performance gap between the two.

Limitations and Future Work

  • Single Previous Frame: Currently, PrevPredMap only leverages predictions from the immediately preceding frame. The authors suggest that incorporating predictions from multiple preceding frames could further enhance performance, though this would introduce significant complexity in post-processing.
  • Map Priors: While the framework is designed to integrate map priors (like outdated HD maps or crowd-sourced data), this aspect is not deeply explored in the current work. Incorporating these priors is seen as a promising avenue for further improving map reliability.

Summary updated

BibTeX
@inproceedings{peng2025prevpredmap,
  author = {Nan Peng and Xun Zhou and Mingming Wang and Xiaojun Yang and Songming Chen and Guisong Chen},
  title = {{PrevPredMap: Exploring Temporal Modeling with Previous Predictions for Online Vectorized HD Map Construction}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2025},
}
Figure