Global Point Cloud Registration Network for Large Transformations

TL;DR

This paper presents ReLaTo (Registration for Large Transformations), an architecture that addresses the cases where large transformations happen while maintaining good performance for local transformations, using a novel Softmax pooling layer to find correspondences in a bilateral consensus manner between two point sets.

Abstract

Three-dimensional data registration is an established yet challenging problem that is key in many different applications, such as mapping the environment for autonomous vehicles, and modeling objects and people for avatar creation, among many others. Registration refers to the process of mapping multiple data into the same coordinate system by means of matching correspondences and transformation estimation. Novel proposals exploit the benefits of deep learning architectures for this purpose, as they learn the best features for the data, providing better matches and hence results. However, the state of the art is usually focused on cases of relatively small transformations, although in certain applications and in a real and practical environment, large transformations are very common. In this paper, we present ReLaTo (Registration for Large Transformations), an architecture that faces the cases where large transformations happen while maintaining good performance for local transformations. This proposal uses a novel Softmax pooling layer to find correspondences in a bilateral consensus manner between two point sets, sampling the most confident matches. These matches are used to estimate a coarse and global registration using weighted Singular Value Decomposition (SVD). A target-guided denoising step is then applied to both the obtained matches and latent features, estimating the final fine registration considering the local geometry. All these steps are carried out following an end-to-end approach, which has been shown to improve 10 state-of-the-art registration methods in two datasets commonly used for this task (ModelNet40 and KITTI), especially in the case of large transformations.

Figures (8)

• Figure 1: Comparison between the error made by the proposed model and the state of the art under large transformations (considering rotations on the $z$ axis) for the KITTI dataset Menze2015. The first image shows the rotation error (RE) and the second is the translation error (TE). Note that the proposed method results in almost no RE. For TE, although one state-of-the-art solution marginally outperforms the proposed model for transformations smaller than 45 degrees, our approach demonstrates superior average performance across the entire range of transformations.
• Figure 2: Proposed architecture for 3D rigid PCR. The two optimized loss functions are marked in red. The proposed novel parts of the architecture are highlighted in yellow. FPS and UPS refer to Farthest Point Sampling and Up-Sampling, respectively. Details are described in Section \ref{['sec:network']}.
• Figure 3: Representation of the target-guided denoising process.
• Figure 4: KITTI dataset registrations. The proposed method successfully registers the point sets, even under large transformations, where other current methods fail most of the time.
• Figure 5: Performance on ModelNet40. Comparison between the proposed model and the state-of-the-art methods under different rotations in the $\mathit{X}$, $\mathit{Y}$, and $\mathit{Z}$ axes. The first row shows the median rotation error (RE) and the second the median translation error (TE).