For human-robot cooperative spray painting robot, offline programming based on predefined model of the unpainted object is a robust and efficient method for trajectory generation. To apply the programmed trajectory on the unpainted object, the relative pose between the object and the predefined model needs to be acquired. Nevertheless, acquiring an accurate estimation of the pose in spray painting setting remains a problem. To address this, a RGB-D pose estimation system based on deep learning and iterative closest point (ICP) alignment is proposed in this paper. The perception module of this system is RGB-D sensor. The RGB-D image of the object is segmented using Fully Convolutional Network (FCN) with RGB-D input. The resulting segmented point cloud is aligned with the model candidates using ICP algorithm to estimate the pose of the object. It is validated in the experiments that the proposed system and methods are effective and robust.