Deep Hough Voting for Robust Global Registration

Abstract

Point cloud registration is the task of estimating the rigid transformation that aligns a pair of point cloud fragments. We present an efficient and robust framework for pairwise registration of real-world 3D scans, leveraging Hough voting in the 6D transformation parameter space. First, deep geometric features are extracted from a point cloud pair to compute putative correspondences. We then construct a set of triplets of correspondences to cast votes on the 6D Hough space, which represents the transformation parameters in the form of sparse tensors. Next, a fully convolutional refinement module is applied to refine the noisy votes. Finally, we identify the consensus among the correspondences from the Hough space, which we use to predict our final transformation parameters. Our method outperforms state-of-the-art methods on the 3DMatch and 3DLoMatch benchmarks, while achieving comparable performance on the KITTI odometry dataset. We further demonstrate the generalizability of our approach by setting a new state of the art on the ICL-NUIM dataset, where we integrate our module into a multi-way registration pipeline.

Proposed Pipeline

Figure 1. Illustration of our proposed module. We compute feature correspondences between point cloud pairs from locally extracted features, which are used to construct a set of triplets. Each triplet votes on the corresponding bin in the sparse Hough space in a parallelizedfashion. We then perform a fully convolutional Hough space refinement to suppress the noisy Hough voting space. Finally, we identify thebin with the max votes, from where we make our final predictions on the transformation parameters.

Efficiency of Proposed Method

Figure 1. (left) Registration recall and the number of registration per second in log scale on 3DMatch benchmark. Our approach is the most accurate overall and fastest among learning-based methods. (right) Analysis on peak memory usage and running time on 3DMatch benchmark

Experimental results

1. Results on 3DMatch & 3DLoMatch


(a) 3DMatch	(b) 3DLoMatch

(c) Precision-recall curve on 3DMatch

Figure 2. (a) Evaluation results on 3DMatch benchmark. The first group of rows shows the results of classical global registration methods, and the second group of rows shows the results of ICP variants. The third group of rows shows the results of learningbased methods. Time includes feature extraction. The reported time of [10] on the original paper was 0.7s, whereas the recalibrated result was 0.96s. (b) Precision-Recall curve on 3DMatch with various threshold values.

2. Results on KITTI & ICL-NUIM


(a) KITTI	(b) ICL-NUIM

Figure 3 . (a) Evaluation results on KITTI test split. All methods are trained on 30cm voxel downsampled point clouds, and thresholds of 0.6m and 5◦ are used to define successful registration. Time includes feature extraction. (b) ATE(cm) on Augmented ICL-NUIM

3. Qualitative results

Figure 4. Pairwise point cloud registration results of DGR [1] (first column) and ours (second column) on 3DMatch dataset [2]. Our method is robust under low overlap and repetitive structures while being about twice as fast as DGR.

Figure 5. Qualitative comparison of DGR [1] (left column) and ours (right column) on KITTI dataset [3].

Acknowlegement

This work was supported by the NRFgrant (2020R1C1C1015260), the IITP grant (No.2019-0-01906, AI Graduate School Program - POSTECH) fundedby Ministry of Science and ICT, Korea, and Samsung Elec-tronics Co., Ltd.

References

Choy, C., Dong, W., and Koltun, V. Deep Global Registration. CVPR, 2020.
Zeng, A., Song, S., Nießner, M., Fisher, M., Xiao, J., and Funkhouser, T. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. CVPR, 2017.
Geiger, A., Lenz, P., and Urtasun, R. Are we ready for autonomous driving? the kitti vision benchmark suite. CVPR, 2012.

Junha Lee	Seungwook Kim	Minsu Cho	Jaesik Park
POSTECH CSE & GSAI