Research map / Video / Video segmentation and tracking
Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation
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 MapUnveiler, a clip-level online vectorized HD map construction model that uses compact clip tokens to unveil occluded map elements within a clip and propagates them across clips via token memory (TTM-style), achieving SOTA on nuScenes and Argoverse2.
Problem
Online vectorized HD map construction from multi-view cameras suffers from occlusions by dynamic objects and degrades over longer perception ranges; existing temporal methods propagate noisy dense BEV features across frames.
Approach
Replace dense spatio-temporal BEV attention with clip-level compact tokens: clip tokens generated from frame-level map queries update BEV features (Intra-clip Unveiler), and memory tokens built with a token summarizer (Inter-clip Unveiler, following TTM) propagate global map information across clips with a temporal stride.
Key results
- On nuScenes 60x30m val, MapUnveiler reaches 68.0 mAP vs 61.5 for MapTRv2 and 66.7 for HIMap, at 12.7 FPS vs 15.6 for MapTRv2.
- Under heavy occlusion splits on nuScenes, mAP is 63.8 vs 53.1 for StreamMapNet and 47.1 for MapTRv2.
- On Argoverse2 60x30m val, mAP 70.5 vs 67.4 for MapTRv2 and 69.6 for HIMap.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add clip-level Intra-clip Unveiler (ablation) | MapTRv2 · nuScenes val, 60 × 30 m | mAP: 61.2% to 67.1% (+5.9%) | Helped |
| add Inter-clip Unveiler for global map relationships (ablation) | MapTRv2 with Intra-clip Unveiler · nuScenes val, 60 × 30 m | mAP: 67.1% to 69.8% (+2.7%) | Helped |
| unfreeze the frame-level MapNet and train end-to-end (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 66.7 to 69.8 | Helped |
| add map queries to the read inputs alongside memory (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 68.9% to 69.8% | Helped |
| write clip tokens to memory (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 67.1%→69.1% | Helped |
| add map tokens to the memory write (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 69.1%→69.8% | Helped |
| add dense BEV features to memory writing (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 69.8 to 68.9 | Hurt |
| increase temporal window from T=3 to T=5 (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 69.8 to 70.1 | Helped |
| use temporal stride S=1 instead of S=2 (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 69.8 to 69.3 | Hurt |
| use temporal stride S=3 instead of S=2 (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 69.8 to 68.9 | Hurt |
| reduce memory token size below 96 (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 69.8 to 68.5 | Hurt |
| increase memory token size above 96 (ablation) | MapUnveiler · nuScenes val, 60 × 30 m | mAP: 69.8 to 69.5; 69.8 to 69.7 | No clear effect |
Surprising findings
- Writing dense BEV features into memory in addition to tokens reduced performance (68.9 vs 69.8 mAP).
- A frozen frame-level MapNet variant still reached SOTA on nuScenes (66.7 mAP).
Limitations
- Peak GPU memory roughly doubles and parameters triple versus MapTRv2 at inference.
- Performance degrades when intermediate frames are dropped (e.g., 20% drop yields 59.6 mAP in one row, with a degraded pattern).
- Fails when regions are invisible across all frames, e.g., boundaries under heavy occlusion are predicted as absent.
Experimental setup
Datasets: nuScenes Argoverse2
Benchmarks: nuScenes Argoverse2
Metrics: AP (Chamfer 0.5/1.0/1.5m averaged) mAP FPS
Baselines: MapTRv2 StreamMapNet SQD-MapNet HIMap MapTR MapQR MGMap MapVR
Builds on: MapTRv2 TTM BEVFormer StreamMapNet
Closest papers by meaning
- Driving by the Rules: A Benchmark for Integrating Traffic Sign Regulations into Vectorized HD Map (CVPR 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 0.87)
- LaneSegNet: Map Learning with Lane Segment Perception for Autonomous Driving (ICLR 2024 · similarity 0.86)
- MGMap: Mask-Guided Learning for Online Vectorized HD Map Construction (CVPR 2024 · similarity 0.86)
- Compact HD Map Construction via Douglas-Peucker Point Transformer (AAAI 2024 · similarity 0.86)
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.