Research map / Point clouds / Registration and matching
PosDiffNet: Positional Neural Diffusion for Point Cloud Registration in a Large Field of View with Perturbations
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 PosDiffNet, a hierarchical window-patch-point point cloud registration model that uses Beltrami graph neural diffusion for feature and position embeddings and a neural-ODE feature-position Transformer, evaluated on Boreas and KITTI under dynamic, weather, and noise perturbations.
Problem
Point cloud registration over large fields of view with dynamic objects, environmental noise (rain, snow, night) and additive noise, where LiDAR sparsity and distortion make learned registration unreliable.
Approach
Combine Beltrami-flow graph neural diffusion that jointly evolves feature and positional embeddings with a Transformer whose attention uses positional embeddings and is integrated as a neural ODE, followed by hierarchical window-patch-point correspondence and SVD/LGR pose estimation.
Key results
- On Boreas (Sunny) trained model, PosDiffNet RTE RMSE 5.73 cm vs. BUFFER 7.49 cm and RoITr 13.05 cm.
- On KITTI 10-m frame setting, PosDiffNet TE 6.6 cm and RE 0.24 deg with RR 99.8%, best among compared methods on TE.
- Ablation: removing the diffusion-based Transformer raises KITTI RTE MAE from 3.97 cm to 11.52 cm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PosDiffNet instead of GeoTrans | Point cloud registration · Boreas (Night), Boreas-trained | RTE MAE (cm): 4.58 -> 4.46 | Helped |
| Use PosDiffNet instead of GeoTrans | Point cloud registration · KITTI, Boreas-trained | RTE MAE (cm): 6.19 -> 4.48 | Helped |
| Use PosDiffNet instead of GeoTrans | Point cloud registration · KITTI, KITTI-trained | RTE MAE (cm): 3.98 -> 4.30 | Hurt |
| Use PosDiffNet instead of GeoTrans | Point cloud registration · Boreas Rain | RTE MAE (cm): 4.96 -> 4.56 | Helped |
| Use PosDiffNet instead of GeoTrans | Point cloud registration · KITTI with Gaussian noise | RTE MAE (cm): 5.37 -> 4.84 | Helped |
| Use PosDiffNet instead of GeoTrans | Point cloud registration · KITTI 10-m frame interval | TE (cm): 6.8 -> 6.6 | Helped |
| Add the window module (ablation) | PosDiffNet · KITTI | RTE MAE (cm): 4.34 -> 3.97 | Helped |
| Add Beltrami diffusion (ablation) | PosDiffNet · KITTI | RTE MAE (cm): 5.20 -> 3.97 | Helped |
| Add the diffusion-based Transformer (ablation) | PosDiffNet · KITTI | RTE MAE (cm): 11.52 -> 3.97 | Helped |
| Replace Geometric Transformer with the diffusion-based Transformer (ablation) | PosDiffNet · KITTI | RTE MAE (cm): 4.23 -> 3.97 | Helped |
| Add Beltrami embedding (ablation) | PosDiffNet Transformer · KITTI | RTE MAE (cm): 4.32 -> 3.97 | Helped |
| Add neural ODE (ablation) | PosDiffNet Transformer · KITTI | RTE MAE (cm): 4.81 -> 3.97 | Helped |
Surprising findings
- On 3DMatch/3DLoMatch indoor benchmarks, PosDiffNet (93.1/76.0 RR) is not best, which the authors attribute to indoor data lacking noise perturbations.
Limitations
- Resource-intensive diffusion and attention computation burdens resource-constrained devices.
- Not best on indoor 3DMatch/3DLoMatch benchmarks.
- Stability analysis is largely theoretical with assumptions such as Lipschitz continuity.
Experimental setup
Datasets: Boreas KITTI synthetic transformations on Boreas and KITTI 3DMatch 3DLoMatch
Benchmarks: Boreas (sunny, night, rain, snow) KITTI KITTI 10-m frame setting 3DMatch 3DLoMatch
Metrics: RTE MAE RTE RMSE RRE MAE RRE RMSE TE RE RR registration recall
Baselines: ICP DCP HGNN++ VCR-Net PCT++ GeoTransformer BUFFER RoITr 3DFeat-Net D3Feat SpinNet Predator
Builds on: Beltrami neural diffusion graph neural PDE GeoTransformer KPConv-FPN EdgeConv (DGCNN) neural ODE
Closest papers by meaning
- Where Precision Meets Efficiency: Transformation Diffusion Model for Point Cloud Registration (AAAI 2025 · similarity 0.88)
- PSReg: Prior-guided Sparse Mixture of Experts for Point Cloud Registration (AAAI 2025 · similarity 0.87)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.87)
- Improved MLP Point Cloud Processing with High-Dimensional Positional Encoding (AAAI 2024 · similarity 0.87)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.86)
- Point Cloud Pre-training with Diffusion Models (CVPR 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.