KdO-Net: Towards Improving the Efficiency of Deep Convolutional Neural Networks Applied in the 3D Pairwise Point Feature Matching

Abstract

In this work, we construct a Kd–Octree hybrid index structure to organize the point cloud and generate patch-based feature descriptors at its leaf nodes. We propose a simple yet effective convolutional neural network, termed KdO-Net, with Kd–Octree based descriptors as input for 3D pairwise point cloud matching. The classic pipeline of 3D point cloud registration involves two steps, viz., the point feature matching and the globally consistent refinement. We focus on the first step that can be further divided into three parts, viz., the key point detection, feature descriptor extraction, and pairwise-point correspondence estimation. In practical applications, the point feature matching is ambiguous and challenging owing to the low overlap of multiple scans, inconsistency of point density, and unstructured properties. To solve these issues, we propose the KdO-Net for 3D pairwise point feature matching and present a novel nearest neighbor searching strategy to address the computation problem. Thereafter, our method is evaluated with respect to an indoor BundleFusion benchmark, and generalized to a challenging outdoor ETH dataset. Further, we have extended our method over our complicated and low-overlapped TUM-lab dataset. The empirical results graphically demonstrate that our method achieves a superior precision and a comparable feature matching recall to the prior state-of-the-art deep learning-based methods, despite the overlap being less than 30 percent. Finally, we implement quantitative and qualitative ablated experiments and visualization interpretations for illustrating the insights and behavior of our network.

Deep-Learning-Based Point Cloud Semantic Segmentation: A Survey

“…For most GCNs, convolution operations are usually only suitable for the feature extraction of structurally fixed graphs. Considering the complexity of graph structures and the heterogeneity in connecting modes, Zhang et al [84] efficiently organized the point cloud by constructing a hybrid index structure based on Kd-Octree and generated patchbased feature descriptors at leaf nodes as input for 3D pairwise point cloud matching. Li et al [85] designed an adaptive graph convolutional neural network, AGCN, which can take arbitrary-sized graphs as input.…”