We explore the effect of geometric structure descriptors on extracting reliable correspondences and obtaining accurate registration for point cloud registration. The point cloud registration task involves the estimation of rigid transformation motion in unorganized point cloud, hence it is crucial to capture the contextual features of the geometric structure in point cloud. Recent coordinates-only methods ignore numerous geometric information in the point cloud which weaken ability to express the global context. We propose Enhanced Geometric Structure Transformer to learn enhanced contextual features of the geometric structure in point cloud and model the structure consistency between point clouds for extracting reliable correspondences, which encodes three explicit enhanced geometric structures and provides significant cues for point cloud registration. More importantly, we report empirical results that Enhanced Geometric Structure Transformer can learn meaningful geometric structure features using none of the following: (i) explicit positional embeddings, (ii) additional feature exchange module such as cross-attention, which can simplify network structure compared with plain Transformer. Extensive experiments on the synthetic dataset and real-world datasets illustrate that our method can achieve competitive results.

Three-dimensional point cloud registration is an important field in computer vision. Recently, due to the increasingly complex scenes and incomplete observations, many partial-overlap registration methods based on overlap estimation have been proposed. These methods heavily rely on the extracted overlapping regions with their performances greatly degraded when the overlapping region extraction underperforms. To solve this problem, we propose a partial-to-partial registration network (RORNet) to find reliable overlapping representations from the partially overlapping point clouds and use these representations for registration. The idea is to select a small number of key points called reliable overlapping representations from the estimated overlapping points, reducing the side effect of overlap estimation errors on registration. Although it may filter out some inliers, the inclusion of outliers has a much bigger influence than the omission of inliers on the registration task. The RORNet is composed of overlapping points’ estimation module and representations’ generation module. Different from the previous methods of direct registration after extraction of overlapping areas, RORNet adds the step of extracting reliable representations before registration, where the proposed similarity matrix downsampling method is used to filter out the points with low similarity and retain reliable representations, and thus reduce the side effects of overlap estimation errors on the registration. Besides, compared with previous similarity-based and score-based overlap estimation methods, we use the dual-branch structure to combine the benefits of both, which is less sensitive to noise. We perform overlap estimation experiments and registration experiments on the ModelNet40 dataset, outdoor large scene dataset KITTI, and natural data Stanford Bunny dataset. The experimental results demonstrate that our method is superior to other partial registration methods. Our code is available at https://github.com/superYuezhang/RORNet .

Learning effective representations from unlabeled data is a challenging task for point cloud understanding. As the human visual system can map concepts learned from 2D images to the 3D world, and inspired by recent multimodal research, we introduce data from point cloud modality and image modality for joint learning. Based on the properties of point clouds and images, we propose CrossNet, a comprehensive intra- and cross-modal contrastive learning method that learns 3D point cloud representations. The proposed method achieves 3D-3D and 3D-2D correspondences of objectives by maximizing the consistency of point clouds and their augmented versions, and with the corresponding rendered images in invariant space. We further distinguish the rendered images into RGB and grayscale images to extract color and geometric features, respectively. These training objectives combine feature correspondences between modalities to combine rich learning signals from point clouds and images. Our CrossNet is simple: we add a feature extraction module and a projection head module to the point cloud and image branches, respectively, to train the backbone network in a self-supervised manner. After the network is pretrained, only the point cloud feature extraction module is required for fine-tuning and directly predicting results for downstream tasks. Our experiments on multiple benchmarks demonstrate improved point cloud classification and segmentation results, and the learned representations can be generalized across domains.

Point cloud registration is a key problem in the application of computer vision to robotics, autopilot and other fields. However, because the object is partially covered up or the resolution of 3D scanners is different, point clouds collected by the same sense may be inconsistent and even incomplete. Inspired by the recently proposed learning-based approaches, we propose Inliers Estimation Network (INENet) which includes a self-designed threshold prediction network and a probability estimation network with adaptive similarity mutual attention to help to find the overlapping area of the point clouds. In order to solve the above problems, we divide the partially overlapping point cloud registration task into two sub-tasks: overlapping areas detection and registration. The threshold prediction network can automatically calculate the threshold according to the input point clouds, and then the probability estimation network estimates the overlapping points by using threshold. The advantages of the proposed approach include: (1) threshold prediction network avoids bias and the complexity of manually adjusting the threshold. (2) Probability estimation network with similarity matrix can deeply fuse the information between a pair of point clouds, which is helpful to improve the accuracy. (3) INENet can be easily integrated into other overlapping region sensitive algorithms and without adjusting parameters. We conduct experiments on the ModelNet40, S3DIS and 3DMatch data sets. Specifically, the rotation error of the registration algorithm integrated with INENet is improved by at least 25% compared with direct partial overlap registration, our method improves the F ₁ score by 5% and has better anti-noise ability compared with the existing overlap detection methods, showing the effectiveness of the proposed method.

Point cloud registration is a critical task in many 3D computer vision studies, aiming to find a rigid transformation that aligns one point cloud with another. In this paper, we propose PANet-a Point-Attention based multi-scale feature fusion network for partially overlapping point cloud registration. This study aims to investigate whether multi-scale features are more effective in improving the precision of alignment compared to fixed-scale local features. PANet comprises two core components: a multi-branch feature extraction module that extracts local features at different scales in parallel, and a Point-Attention Module that learns an appropriate weight for each branch and then fuse these multi-scale features by weighted combination to enhance the representation ability of features. At the end of the network, four hidden layers are used to obtain the rigid transformation from the source point cloud to the template point cloud. Experiments on the synthetic ModelNet40 dataset demonstrate that PANet outperforms state-of-the-art performance in terms of both alignment precision and robustness against noise. PANet also exhibits strong generalization ability on real-world Stanford 3D and ICL-NUIM datasets. In addition, the computational complexity of our model compared to previous works is also evaluated. The results and ablation studies demonstrate that multi-scale fused local features are better at improving registration accuracy than fixed-scale local features. The findings may inspire future research in related fields and contribute to the development of new ideas and approaches.

Extracting discriminative representations is the key step for correspondence-free point cloud registration. The extracted representations require to be discriminative to transformation, which demands representations to reduce the influence of redundant information irrelevant to transformation. However, recently proposed methods ignore this crucial property, resulting in limited ability to represent point cloud. In addition, researching correspondence-free point cloud registration has stagnated in recent years. In this paper, we try to relieve features redundancy issue for correspondence-free point cloud registration from a new perspective. Specifically, our method comprises two stages: feature extraction stage and rigid body transformation stage. In feature extraction stage, we aim to maximize multi-hierarchical mutual information between different hierarchical features, which can provide discriminative and less redundancy representations to regress transformation parameters for next stage. In rigid body transformation stage, we utilize dual quaternion to estimate transformation parameters, which combines rotation and translation simultaneously within a unified framework and obtains a compact representations for rigid transformation. The proposed model is trained in an unsupervised manner on the ModelNet40 dataset. The experimental results illustrate that our method achieves higher accuracy and robustness compared with existing correspondence-free methods.