Research map / Label-efficient and robust learning / Recognition and long-tail

Leveraging SD Map to Augment HD Map-based Trajectory Prediction

CVPR 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SATP, which augments HD map-based trajectory prediction models with OpenStreetMap SD maps via a parallel SD-HD query fusion and an AlignNet that learns to align SD to HD maps, improving HiVT, DenseTNT, QCNet and VAD.

Problem

Online HD maps used by trajectory prediction models suffer from perception errors and feature redundancy, which degrade prediction performance in deployed autonomous driving systems.

Approach

Abstracts how HD-map-based predictors encode maps and extract features, builds downscaled SD-map counterparts of those modules, adaptively fuses HD and SD queries before the prediction head, and uses a 2D point transformer AlignNet that predicts a relative pose for SD-to-HD alignment, trained only through the trajectory loss.

Key results

• SATP improves HiVT, DenseTNT and QCNet with online HD maps by up to 25% (QCNet with HRMapNet: MR 0.062 vs 0.083), with average gains of 9% ADE, 13% FDE and 17% MR.
• On Argoverse2 with simulated noisy HD maps, QCNet with SATP reaches ADE 0.752 and MR 0.164, versus 0.884 and 0.206 without SATP.
• SATP adds 0.2-0.4M parameters and 6-8% inference latency to the baselines.

Measured effects reported in the paper

Surprising findings

• Using SD maps to enhance online HD map construction (P-MapNet) gives smaller trajectory gains than feeding SD maps directly into the predictor (SATP).
• Sequential SD-HD fusion orders perform worse than parallel adaptive fusion.

Limitations

• Does not model delayed updates of SD maps.
• Uses simulated perception errors on Argoverse2 rather than real online HD maps.
• SD map positioning noise is simulated with Gaussian noise rather than real localization errors.

Experimental setup

Datasets: nuScenes Argoverse2 OpenStreetMap

Benchmarks: nuScenes Argoverse2

Metrics: minADE6 minFDE6 MR6 mAP L2 displacement error collision rate parameter count latency

Baselines: HiVT DenseTNT QCNet MapTRv2 HRMapNet P-MapNet VAD-tiny SD-based trajectory prediction (Ref 22)

Builds on: HiVT DenseTNT QCNet MapTRv2 Point Transformer

Closest papers by meaning

• Producing and Leveraging Online Map Uncertainty in Trajectory Prediction (CVPR 2024 · similarity 0.89)
• SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.89)
• MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.88)
• InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.85)
• Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map (CVPR 2025 · similarity 0.85)
• The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset (NeurIPS 2026 · similarity 0.85)