摘要
我们提出了一种名为“深度全局配准”（Deep Global Registration）的可微分框架，用于真实世界三维扫描数据的两两配准。该框架由三个模块构成：用于预测对应关系置信度的六维卷积网络、用于闭式位姿估计的可微分加权普鲁克斯特（Weighted Procrustes）算法，以及用于位姿精化的鲁棒梯度优化SE(3)优化器。实验结果表明，该方法在真实世界数据上的表现优于当前最先进的基于学习的方法及传统方法。
代码仓库
chrischoy/DeepGlobalRegistration
官方
pytorch
chrischoy/FCGF
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| point-cloud-registration-on-3dlomatch-10-30 | DGR (reported in REGTR) | Recall ( correspondence RMSE below 0.2): 48.7 |
| point-cloud-registration-on-3dmatch-at-least-1 | DGR (RE (all), TE(all) are reported in PCAM) | RE (all): 9.5 Recall (0.3m, 15 degrees): 91.3 TE (all): 0.25 |
| point-cloud-registration-on-3dmatch-at-least-2 | DGR (reported in REGTR) | Recall ( correspondence RMSE below 0.2): 85.3 |
| point-cloud-registration-on-kitti-fcgf | DGR (RE (all), TE(all) are reported in PCAM) | RE (all): 1.62 Recall (0.6m, 5 degrees): 96.9 TE (all): 0.34 |
| point-cloud-registration-on-kitti-fcgf | DGR + ICP (RE (all), TE(all) are reported in PCAM) | RE (all): 1.43 Recall (0.6m, 5 degrees): 98.2 TE (all): 0.16 |