Research map / Point clouds / Point cloud classification
Compact HD Map Construction via Douglas-Peucker Point Transformer
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 DPFormer, a DETR-like transformer that represents HD map elements with compact Douglas-Peucker points plus auxiliary points, trained with uncertainty-aware position loss and pairwise-point shape matching loss.
Problem
Online HD map construction suffers from redundant point representations (equidistant points) or complex curve modeling (Bezier), and ambiguous point regression in BEV perception.
Approach
A class-agnostic element representation using Douglas-Peucker points (sparse, at structurally important locations) combined with auxiliary points, supervised by hierarchical constraints: instance, point-level classification, Laplace-based position-with-uncertainty loss, and pairwise-point Gaussian KL shape matching.
Key results
- DPFormer achieves 60.8 mAP on nuScenes with 150 epochs, versus 58.7 for MapTR (110 epochs) under the same single-GPU setup comparison in Table 1.
- DPFormer uses a compression ratio of 46.7% versus 205.1% for MapTR, roughly 4x fewer points.
- Under 2 m and 5 m neighborhoods defined by GT DP points, DPFormer improves mAP over MapTR by 7.4 and 5.8 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DP points instead of equidistant points (ablation) | DPFormer · nuScenes val | mAP: 51.6 -> 52.2 | Helped |
| Add position loss with uncertainty instead of point2point loss (ablation) | DPFormer with DP points · nuScenes val | mAP: 52.2 -> 53.3 | Helped |
| Add pairwise-point shape matching loss instead of cosine similarity loss (ablation) | DPFormer with position loss with uncertainty · nuScenes val | mAP: 53.3 -> 54.6 | Helped |
| Set position-loss weight αp to 5.0 rather than 4.0 (ablation) | DPFormer · nuScenes val | mAP: 52.3 -> 54.6 | Helped |
| Increase position-loss weight αp from 5.0 to 6.0 (ablation) | DPFormer · nuScenes val | mAP: 54.6 -> 53.1 | Hurt |
| Add pairwise-point shape matching constraint (ablation) | DPFormer · nuScenes val | mAP: 51.3 -> 54.6 | Helped |
| Increase pairwise-point shape matching weight αs from 0.5 to 0.75 (ablation) | DPFormer · nuScenes val | mAP: 54.6 -> 52.7 | Hurt |
| Use DPFormer instead of MapTR | R50 map construction · nuScenes val | mAP: +2.1% | Helped |
Surprising findings
- Over half of map elements (52.8%) can be represented accurately with only 2 DP points, reflecting that most elements are straight lines.
Limitations
- Ground-truth DP point generation sacrifices accuracy, limiting gains on boundaries with complex topology.
- Pairwise shape loss weight too large causes position ambiguity (Table 5).
Experimental setup
Datasets: nuScenes
Benchmarks: nuScenes
Metrics: AP mAP Chamfer distance compression ratio FPS
Baselines: HDMapNet VectorMapNet MapTR
Builds on: MapTR GKT Hungarian matching Focal loss DSANet
Closest papers by meaning
- HIMap: HybrId Representation Learning for End-to-end Vectorized HD Map Construction (CVPR 2024 · similarity 0.86)
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.86)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.85)
- MGMap: Mask-Guided Learning for Online Vectorized HD Map Construction (CVPR 2024 · similarity 0.85)
- Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction (CVPR 2025 · similarity 0.84)
- InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.