Li, HuajianHuajianLiKraljevski, IvanIvanKraljevskiMeyer, PaulPaulMeyerTschöpe, ConstanzeConstanzeTschöpeWolff, MatthiasMatthiasWolff2025-01-172025-01-172024-12-20https://publica.fraunhofer.de/handle/publica/48142710.1109/SENSORS60989.2024.10784539In this paper, we present a novel deep learning-integrated pipeline called YOLO-ICP that aims to estimate the six degree of freedom (6-DoF) pose of objects using RGB-D sensors and does not require pose labels to train deep learning networks. YOLO-ICP integrates a real-time object detection algorithm with a point cloud registration method to estimate the pose of multiple objects. We evaluated our approach by quantitatively comparing it with baseline models on the OccludedLINEMOD dataset. Experimental results illustrate that our approach outperforms baseline models in challenging scenarios with textureless and occluded objects. In particular, our pipeline shows superior performance when dealing with small and symmetric objects in terms of the ADD(-S) metric.enPoint cloudDeep learningYOLOReal-time systemsPose estimation6-DOFSensorsCADbin-pickingRGB-D camera600 Technik, Medizin, angewandte Wissenschaftenconference paper