2-D Deep Learning Model on 3-D Image Segmentation

For image segmentation of 3D objects, researchers recently focus on 3D image deep learning methods. These studies aim at developing robust and accurate deep learning models. However, 3D deep learning methods using point-cloud data are time-consuming. In this paper, we present a novel 3D image segmentation method based on a 2D deep learning model to achieve efficient segmentation performance. While using a single camera angle to collect object 3D information, we merely use the depth map and adopt a 2D deep-learning model to segment objects on the scene. We validated the proposed method by comparison with Point-Net. In our experiment, we provided an extensive comparison between Point-Net and our 2D deep-learning model. The result shows that our model had a similar accuracy but is much faster than Point-Net.