Christopher Choy, Wei Dong, Vladlen Koltun
We present Deep Global Registration, a differentiable framework for pairwise registration of real-world 3D scans. Deep global registration is based on three modules: a 6-dimensional convolutional network for correspondence confidence prediction, a differentiable Weighted Procrustes algorithm for closed-form pose estimation, and a robust gradient-based SE(3) optimizer for pose refinement. Experiments demonstrate that our approach outperforms state-of-the-art methods, both learning-based and classical, on real-world data.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Point Cloud Registration | 3DLoMatch (10-30% overlap) | Recall ( correspondence RMSE below 0.2) | 48.7 | DGR (reported in REGTR) |
| Point Cloud Registration | 3DMatch (at least 30% overlapped - FCGF setting) | RE (all) | 9.5 | DGR (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | 3DMatch (at least 30% overlapped - FCGF setting) | Recall (0.3m, 15 degrees) | 91.3 | DGR (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | 3DMatch (at least 30% overlapped - FCGF setting) | TE (all) | 0.25 | DGR (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | KITTI (FCGF setting) | RE (all) | 1.43 | DGR + ICP (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | KITTI (FCGF setting) | Recall (0.6m, 5 degrees) | 98.2 | DGR + ICP (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | KITTI (FCGF setting) | TE (all) | 0.16 | DGR + ICP (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | KITTI (FCGF setting) | RE (all) | 1.62 | DGR (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | KITTI (FCGF setting) | Recall (0.6m, 5 degrees) | 96.9 | DGR (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | KITTI (FCGF setting) | TE (all) | 0.34 | DGR (RE (all), TE(all) are reported in PCAM) |
| Point Cloud Registration | 3DMatch (at least 30% overlapped - sample 5k interest points) | Recall ( correspondence RMSE below 0.2) | 85.3 | DGR (reported in REGTR) |
| 3D Point Cloud Interpolation | 3DLoMatch (10-30% overlap) | Recall ( correspondence RMSE below 0.2) | 48.7 | DGR (reported in REGTR) |
| 3D Point Cloud Interpolation | 3DMatch (at least 30% overlapped - FCGF setting) | RE (all) | 9.5 | DGR (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | 3DMatch (at least 30% overlapped - FCGF setting) | Recall (0.3m, 15 degrees) | 91.3 | DGR (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | 3DMatch (at least 30% overlapped - FCGF setting) | TE (all) | 0.25 | DGR (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | KITTI (FCGF setting) | RE (all) | 1.43 | DGR + ICP (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | KITTI (FCGF setting) | Recall (0.6m, 5 degrees) | 98.2 | DGR + ICP (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | KITTI (FCGF setting) | TE (all) | 0.16 | DGR + ICP (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | KITTI (FCGF setting) | RE (all) | 1.62 | DGR (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | KITTI (FCGF setting) | Recall (0.6m, 5 degrees) | 96.9 | DGR (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | KITTI (FCGF setting) | TE (all) | 0.34 | DGR (RE (all), TE(all) are reported in PCAM) |
| 3D Point Cloud Interpolation | 3DMatch (at least 30% overlapped - sample 5k interest points) | Recall ( correspondence RMSE below 0.2) | 85.3 | DGR (reported in REGTR) |