Summary
2023
Session Number:C2L-1
Session:
Number:C2L-11
Scale-Equivariant Convolution for Projection-Based Point Cloud Segmentation
Marumo Hidetaka, Matsubara Takashi,
pp.443-446
Publication Date:2023-09-21
Online ISSN:2188-5079
Summary:
With the progress of Artificial intelligence (AI) computer vision, semantic segmentation of Light Detection and Ranging (LiDAR) point clouds using deep neural networks (DNNs) has attracted attention in autonomous driving. Considering recognition accuracy and computational complexity, a promising approach is to project a point cloud into a 2-dimensional (2D) range image and process it with 2D convolutional neural networks (CNNs). Since distant objects appear smaller than nearby objects in an image, it is crucial to incorporate scale-equivariance into the CNN to improve parameter efficiency and recognition accuracy, but no method focuses on it. We propose a new scale-equivariant convolution method, focusing on the relationship between object distance and scale ratio in images and the theoretical properties of partial differential operators. Evaluation experiments on the LiDAR point cloud dataset demonstrate the effectiveness of our method.