Spatial and Semantic Information Enhancement for Indoor 3D Object Detection

Abstract

Object detection technology is one of the key technologies for indoor service robots. However, due to the various types of objects in the indoor environment, the mutual occlusion between the objects is serious, which increases the difficulty of object detection. In view of the difficult challenges of object detection in the indoor environment, we propose an indoor three-dimensional object detection based on deep learning. Most existing 3D object detection techniques based on deep learning lack sufficient spatial and semantic information. To address this issue, the article presents an indoor 3D object detection method with enhanced spatial semantic information. This article proposes a new (Edge Convolution+) EdgeConv+, and based on it, a Shallow Spatial Information Enhancement module (SSIE) is added to Votenet. At the same time, a new attention mechanism, Convolutional Gated Non-Local+ (CGNL+), is designed to add Deep Semantic Information Enhancement module (DSIE) to Votenet. Experiments show that on the ScanNet dataset, the proposed method is 2.4% and 2.1% higher than Votenet at mAP@0.25 and mAP@0.5, respectively. Furthermore, it has strong robustness to deal with sparse point clouds