Fast 3D Object Detection with RGB-D Images Using Graph Convolutional Network

Abstract

In this paper, a lightweight 3D object detection model using color and depth images is proposed. In recent years, several studies have focused on the application of deep learning to object detection. They use many techniques, including improved feature extraction methods and instance segmentation, to increase the accuracy. However, such 2D object detection has its limitations. Other models and methods are needed to deal with occlusion and to identify 3D positions. In contrast, there have been many studies in this field applying deep learning to 3D object detection. However, many of them are computationally expensive and difficult to run in real time because they deal with dense point clouds. In the proposed model, after feature extraction from the color image, a sparse point cloud is created from the range image to achieve fast object detection. Graph convolution for point clouds and feature extraction with depth information are also used. As a result, the proposed model achieved 56.4 fps when using ResNet34.