Research map / 3D vision / Depth and stereo
MapTracker: Tracking with Strided Memory Fusion for Consistent Vector HD Mapping
Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, Yasutaka Furukawa
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add memory buffer with fusion of past latents (latest 4 frames) to tracking (ablation) | MapTracker · nuScenes consistent GT | mAP: 70.8 -> 74.9 | Improved |
| add distance-strided selection of memory frames (ablation) | MapTracker · nuScenes consistent GT | mAP: 74.9 -> 76.1 | Improved |
| add distance-strided selection of memory frames (ablation) | MapTracker · nuScenes consistent GT | C-mAP: 68.1 -> 69.1 | Improved |
| formulate mapping as tracking with single embedding memory instead of conditional detection (ablation) | StreamMapNet · nuScenes consistent GT | C-mAP: 56.4 -> 62.4 | Improved |
| use larger strides {5,10,15,20} m without a near frame (ablation) | MapTracker · nuScenes old split 60x30m | C-mAP: 69.1 -> 51.5 | Hurt |
| use small strides {1,2,3,4} m instead of {1,5,10,15} m (ablation) | MapTracker · nuScenes old split 60x30m | mAP: 76.1 -> 75.8 | Hurt |
| buffer of 20 with zero strides vs buffer of 4 latest frames (ablation) | MapTracker · nuScenes old split 60x30m | mAP: 74.9 -> 75.0 | No clear effect |
| increase track-extraction look-back from 1 to 5 frames for baselines (ablation) | MapTRv2 · nuScenes | C-mAP: 50.5 -> 60.0 | Improved |
| replace StreamMapNet with MapTracker on geographically non-overlapping splits | StreamMapNet · nuScenes 100x50m geo-disjoint | mAP: 23.0 -> 36.2 | Improved |
| add memory fusion layers and tracking query propagation relative to StreamMapNet | MapTracker · nuScenes inference speed | FPS: 14.2 -> 11.5 | Hurt |
Closest papers by meaning
- Unveiling the Hidden: Online Vectorized HD Map Construction with Clip-Level Token Interaction and Propagation (NeurIPS 2024 · similarity 0.88)
- VectorMapNet: End-to-end Vectorized HD Map Learning (ICML 2023 · similarity 0.88)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.86)
- Patched Line Segment Learning for Vector Road Mapping (AAAI 2024 · similarity 0.85)
- InteractionMap: Improving Online Vectorized HDMap Construction with Interaction (CVPR 2025 · similarity 0.85)
- SpatialTracker: Tracking Any 2D Pixels in 3D Space (CVPR 2024 · similarity 0.85)
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.
