PredMapNet

Future and Historical Reasoning for Consistent Online HD Vectorized Map Construction

Bo Lang*, Nirav Savaliya, Zhihao Zheng*, Jinglun Feng, Zheng-Hang Yeh, Mooi Choo Chuah*

Honda Research Institute USA, * Lehigh University

📄 Paper    💻Code

Framework Overview

Overview

PredMapNet is an end-to-end framework for consistent online HD vectorized map construction designed for autonomous driving systems.

Unlike prior query-based methods that rely on random initialization and implicit temporal modeling, PredMapNet introduces explicit historical memory and short-term future reasoning to produce temporally stable and coherent map predictions across frames.

The framework improves both spatial alignment and temporal consistency, enabling reliable online vectorized HD mapping.

Results

Motivation

Watch the Video

High-definition (HD) vectorized maps are critical for:

  • Autonomous navigation
  • Motion planning
  • Scene understanding

However, existing query-based map construction methods often:

  • Use randomly initialized queries
  • Lack explicit temporal modeling
  • Produce temporally inconsistent predictions
  • Struggle with stable instance tracking over time

PredMapNet addresses these limitations by integrating semantic-aware query generation, history map memory, and future motion priors into a unified framework.

Key Contributions

1: Semantic-Aware Query Generator (SAQG)

  • Replaces random query initialization with semantically grounded queries
  • Aligns queries with scene context using semantic masks
  • Improves spatial precision and instance localization

2: History Rasterized Map Memory

  • Maintains fine-grained instance-level rasterized maps from previous frames
  • Provides explicit historical priors
  • Supports stable long-term instance tracking

3: History-Map Guidance Module (HMG)

  • Injects historical map memory into the decoder
  • Guides query refinement using past instance information
  • Enhances temporal continuity across frames

4: Short-Term Future Guidance (STFG)

  • Predicts short-term future locations of map instances
  • Incorporates motion priors from historical trajectories
  • Further stabilizes predictions in dynamic scenes

5: State-of-the-Art Performance

Results Old Split

Results New Split

Links

Paper: https://arxiv.org/pdf/2602.16669

Code: Coming Soon

Cite our work:

If you find this work useful, please consider citing:

@article{predmapnet2026,
  title={PredMapNet: Future and Historical Reasoning for Consistent Online HD Vectorized Map Construction},
  author={Lang, Bo and Savaliya, Nirav and Zheng, Zhihao and Feng, Jinglun and Yeh, Zheng-Hang and Chuah, Mooi Choo},
  booktitle = {Winter Conference on Applications of Computer Vision (WACV), 2026},
  year={2026}
}