Monocular Image-Based 3D Object Detection for Industrial Pick-and-Place Tasks

This paper proposes a new pipeline that estimates depth information from a single RGB image and utilizes it for object detection. By fusing the pseudo depth map generated through the MiDaS-based Monocular Depth Estimation with the RGB image, the performance of the existing RGB-based detection model is improved. The proposed method is practical because spatial structure information can be utilized without an actual depth sensor, and shows robust performance, especially in occlusion between objects and complex background environments. As a result of experiments on the KITTI and SUN RGB-D datasets, the proposed method achieved performance improvement of 4.2% and 3.8% on mAP compared to the existing RGB-based method, respectively.