Research map / Point clouds / Registration and matching
GraphI2P: Image-to-Point Cloud Registration with Exploring Pattern of Correspondence via Graph Learning
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 GraphI2P, an image-to-point cloud registration method that bridges modalities with a virtual point cloud, decouples geometry in a virtual-spherical representation, adaptively samples via a log-normal MLE distribution, and prunes false 2D-3D correspondences with graph learning, evaluated on KITTI and nuScenes.
Problem
Image-to-point cloud registration fails when the image and LiDAR scan come from non-synchronized frames, due to the modality gap between images and LiDAR and the domain gap between virtual and LiDAR point clouds, plus errors from frustum filtering.
Approach
Replaces frustum selection with graph-based global correspondence selection that uses correspondence feature-distance graphs and a GNN to prune false matches; adds a virtual-spherical representation that decouples direction from range, and a log-normal MLE-based distribution-adaptive sampler to align virtual and LiDAR point cloud distributions.
Key results
- On non-synchronized random-frame setting, GraphI2P reaches 95.40% RR on KITTI (RTE 0.68 m, RRE 1.91°), claimed 10.77% above CFI2P.
- On nuScenes random-frame setting, RR 93.58% versus CFI2P 81.05%, claimed 12.53% gain.
- Same-frame KITTI RR 99.61% vs CFI2P 97.15%; ablation shows removing graph selection drops RR from 95.40% to 80.25%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GraphI2P instead of CFI2P | Image-to-point-cloud registration · KITTI Odometry, non-synchronized random-frame | RR (%): 84.63 -> 95.40; +10.77% | Helped |
| Use GraphI2P instead of CFI2P | Image-to-point-cloud registration · nuScenes, non-synchronized random-frame | RR (%): 81.05 -> 93.58; +12.53% | Helped |
| Add virtual-spherical representation (ablation) | GraphI2P · KITTI Odometry, non-synchronized random-frame | RR (%): 90.07 -> 95.40 | Helped |
| Add adaptive distribution sampling (ablation) | GraphI2P · KITTI Odometry, non-synchronized random-frame | RR (%): 84.81 -> 95.40 | Helped |
| Add graph-based correspondence selection (ablation) | GraphI2P · KITTI Odometry, non-synchronized random-frame | RR (%): 80.25 -> 95.40 | Helped |
| Reduce input point cloud density from 40960 to 5120 points (ablation) | GraphI2P · KITTI Odometry, non-synchronized random-frame | RR (%): 95.40 -> 75.81 | Hurt |
| Reduce input point cloud density from 40960 to 10960 points (ablation) | GraphI2P · KITTI Odometry, non-synchronized random-frame | RR (%): 95.40 -> 87.85 | Hurt |
| Reduce input point cloud density from 40960 to 20480 points (ablation) | GraphI2P · KITTI Odometry, non-synchronized random-frame | RR (%): 95.40 -> 93.64 | Hurt |
Surprising findings
- Classic point-to-point registration methods (GeoTransformer, Hunter) applied to virtual point clouds perform poorly, attributed to domain and distribution gap.
Limitations
- Evaluated only on yaw-only (2D) random perturbations; 3D random rotation not yet tested.
- Performance degrades significantly when point cloud density falls below a threshold (e.g., 5120 points).
Experimental setup
Datasets: KITTI Odometry nuScenes
Benchmarks: KITTI Odometry nuScenes
Metrics: RTE RRE registration recall inlier ratio
Baselines: CFI2P VP2P CorrI2P DeepI2P FreeReg RetrI2P 2D3D-Matr GeoTransformer Hunter
Builds on: FreeReg KPConv circle loss ZoeDepth VP2P
Closest papers by meaning
- Robust Point Cloud Registration Framework Based on Deep Graph Matching (CVPR 2021 · similarity 0.91)
- A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration (NeurIPS 2024 · similarity 0.88)
- SG-PGM: Partial Graph Matching Network with Semantic Geometric Fusion for 3D Scene Graph Alignment and Its Downstream Tasks (CVPR 2024 · similarity 0.88)
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.87)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 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.