Robust local feature descriptor for 3D point cloud analysis using siamese networks

Abstract

Point cloud processing, as a growing research domain in machine vision, has gained increasing importance with the advancement of low-cost 3D sensors. Utilizing 3D point cloud data for tasks like 3D object detection and robotic localization entails challenges such as occlusion and clutter. This paper proposes a novel local feature descriptor leveraging deep learning techniques to extract precise and efficient geometric information from the surrounding environment of a point. By integrating T-Net and Siamese networks, the proposed method is capable of extracting consistent local features for points that are rigidly equivalent, thus providing high accuracy in matching 3D objects under challenging conditions. Unlike recent attention- and diffusion-based descriptors, which achieve strong accuracy at the cost of increased computational complexity, the proposed method achieves a favorable trade-off between matching accuracy and computational efficiency: it operates with linear, O(N) complexity in the number of input points and a lightweight architecture of approximately 0.75 million parameters, making it well-suited for real-time and resource-constrained applications. The method not only demonstrates outstanding performance in addressing existing challenges in 3D point cloud processing, such as noise and occlusion, but also exhibits a strong ability to detect and precisely match 3D objects. Evaluation results using the Bologna and 3DMatch datasets show that the proposed method achieves competitive precision, recall, and fragment matching recall compared to existing methods, achieving an average fragment matching recall of 91.23%. The proposed approach has potential applications in various fields, including machine vision, image processing, and augmented reality.