F-3DNet: Leveraging Inner Order of Point Clouds for 3D Object Detection

Abstract

Point clouds are often perceived as irregular and disorderly data in Internet of Things (IoT) applications. However, these point clouds possess implicit order and context information due to the laser arrangement and sequential scanning process, which are often overlooked. In this paper, we propose a novel method called Frustum 3DNet (F-3DNet) for 3D object detection from point clouds in IoT. Our approach utilizes the inner order of point clouds to construct a rearranged feature matrix and generate a pseudo panorama from LiDAR data. Based on the pseudo image, we extend 2D region proposals to 3D space and obtain frustum regions of interest. For each frustum, we generate a sequence of small frustums by slicing over distance, and introduce a novel local context feature extraction module to incorporate context information. The extracted context features are then concatenated with frustum features and fed to a fully convolutional network (FCN), followed by a classifier and a regressor. We further refine and fuse the output with RGB input to improve the outcome. Ablation studies verify the effectiveness of our proposed components. Experimental results on KITTI and nuScenes datasets demonstrate that F-3DNet outperforms existing methods in IoT.