Research map / Detection and segmentation / 3D detection and driving

MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert

AAAI 2025 · Detection and segmentation › 3D detection and driving · approach: LiDAR perception for driving · 5 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 MapExpert, an online HD map construction method using sparse map-element experts (top-K routing with SwiGLU experts), a balance loss, and a Learnable Weighted Moving Descent temporal fusion module, reporting state-of-the-art mAP on nuScenes and Argoverse2.

Problem

Online HD map construction from surround cameras struggles because non-cubic map elements (Bézier-like lane dividers, slender erratic road boundaries, rectangular crossings) are poorly fit by unified DETR-like decoders, and stacking historical BEV features lets outdated history dominate the current

Approach

Replace the unified decoder with a sparse mixture-of-experts transformer layer whose routers send map-element queries to top-K SwiGLU experts, add an expert load-balance loss, and fuse a single previous fused BEV feature with a learnable weight on the current feature (LWMD) instead of stacking multiple frames.

Key results

• On nuScenes original split (50m, ResNet50, 100 epochs), MapExpert reaches 76.5 mAP, 1.8 above the best prior method (MapTracker at 74.7 mAP in the table's comparison).
• On original Argoverse2 split (30m), MapExpert reaches 72.8 mAP, 1.4 above MapTracker (71.4).
• Ablation on nuScenes: the expert layer adds about 1.1 mAP over baseline, the expert balance loss pushes the total to 76.2 mAP, and LWMD plus slice branch bring it to 76.5.

Measured effects reported in the paper

Surprising findings

• Adding the sparse expert layer alone gives only about 0.4% mAP gain over baseline in the component ablation, while the balance loss contributes most of the later gain.
• Robustness experiments show that router input jitter with std around 0.05 or more sharply degrades performance (mAP falls from 39.7 to 25.6 at ratio 0.05 in 20-epoch runs).

Limitations

• All six inactive experts still consume memory during inference, since only 2 of 8 experts are active.
• Training is unstable and needs a large batch size and an auxiliary segmentation loss to converge.
• The encoder is resource-intensive, and the method is not shown to run in real time.

Experimental setup

Datasets: nuScenes Argoverse2

Benchmarks: nuScenes Argoverse2

Metrics: average precision (AP) mean average precision (mAP)

Baselines: HDMapNet VectorMapNet MapTR MapTRv2 MapQR StreamMapNet MapTracker

Builds on: MapTracker Mixture of Experts BEVFormer StreamMapNet

Closest papers by meaning

• Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.88)
• SDTagNet: Leveraging Text-Annotated Navigation Maps for Online HD Map Construction (NeurIPS 2025 · similarity 0.88)
• Leveraging SD Map to Augment HD Map-based Trajectory Prediction (CVPR 2025 · similarity 0.88)
• InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.88)
• Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction (CVPR 2025 · similarity 0.87)
• Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map (CVPR 2025 · similarity 0.87)