DMSC-Net: A deep Multi-Scale context network for 3D object detection of indoor point clouds

Indoor object detection has emerged as one of the key technologies for the success of numerous indoor system applications, such as autonomous navigation, accurate modeling of indoor environments, digital twin and terra Hertz (THz) communications. This paper first proposes a flexible and inter-operational detection module, termed deep multi-scale context (DMSC) module, aiming at the development of efficient indoor object detection techniques using the point clouds. More specifically, by combining the deep contextual information of indoor objects and multi-scale features, a novel deep multi-scale contextual feature is designed. Furthermore, we introduce the decoder part of the vision transformer into the indoor object proposal generation by means of a multi-head attention (MHA) module from a three-dimensional (3D) point cloud to accurately extract object proposals generating high-quality bounding boxes. Extensive experiments have shown that, the effective interoperability of the proposed DMSC module with three object detection networks, namely VoteNet, GroupFree 3D and RBGNet, leads to improvements in their mAP@0.25 by 6.5%, 0.9% and 0.4% on the ScanNetV2 datasets, respectively. The proposed end-to-end network, termed as DMSC-Net, consists of an indoor point cloud feature learning backbone (FLB) unit, and three modules, namely the DMSC, a voting decision (VD) module, and an MHA module. Extensive experiments have shown that the DMSC-Net outperforms other advanced indoor 3D detection networks, such as RBGNet, by 1.1% and 0.9% of mAP@0.25 when applied on ScanNet and SUN RGB-D datasets, respectively. The developed code is publicly available at: https://github.com/CNU-DLandCV-lab/MHA_DMSC.