[4C1-GS-7] Vision, speech media processing

[4C1-GS-7-01] Range-Equivariant Convolution for Semantic Segmentation of LiDAR Point Clouds

〇Hidetaka Marumo¹, Takashi Matsubara¹ (1. Osaka university)

Keywords:LiDAR point clouds, Scale equivariance, autonomous driving

In autonomous driving, semantic segmentation of LiDAR point clouds has attracted much attention, and various methods have been proposed. For efficiency and ease of design, the mainstream methods convert the point clouds into 2D range-images by spherical projection and feed them to 2D convolutional neural networks. To boost the accuracy, scale-equivariance incorporated into the network is crucial because distant objects are smaller than nearby ones in images. However, to our best knowledge, no method has focused on scale-equivariance. In this paper, we focus on the relationship between the object distance and the scale ratio in images and propose a novel scale-equivariant convolutional method. The kernels in this method are defined as linear combinations of partial differential operators (PDO), and scaled features are transformed into unscaled ones by weighting kernels according to the distance of objects and differential order of corresponding PDOs. We tested the effectiveness of REconv by replacing the standard convolution in the encoder of RangeNet21 with REconv. Our experiments were conducted on the Semantic KITTI dataset, and mIoU was improved by 0.5% from baseline on the validation set. This result showed that REconv is effective for semantic segmentation of LiDAR point clouds.