Abstract
Generating 3D point clouds from a single image has attracted full attention from researchers in the field of multimedia, remote sensing and computer vision. With the recent proliferation of deep learning, various deep models have been proposed for the 3D point cloud generation. However, they require objects to be captured with absolutely clean backgrounds and fixed viewpoints, which highly limits their application in the real environment. To guide 3D point cloud generation, we propose a novel network, RealPoint3D, to integrate prior 3D shape knowledge into the network. Taking additional 3D information, RealPoint3D can handle 3D object generation from a single real image captured from any viewpoint and complex background. Specifically, provided a query image, we retrieve the nearest shape model from a pre-prepared 3D model database. Then, the image, together with the retrieved shape model, is fed into RealPoint3D to generate a fine-grained 3D point cloud. We evaluated the proposed RealPoint3D on the ShapeNet dataset and ObjectNet3D dataset for the 3D point cloud generation. Experimental results and comparisons with state-of-the-art methods demonstrate that our framework achieves superior performance. Furthermore, our proposed framework works well for real images in complex backgrounds (the image has the remaining objects in addition to the reconstructed object, and the reconstructed object may be occluded or truncated) with various viewing angles.
Highlights
Generating 3D point clouds from a single image, which aims to provide solutions to tasks such as autonomous driving, virtual reality and robotic surgery, is a fundamental, intriguing area in remote sensing and computer vision
We compared our approach with the Point Set Generation Network (PSGN) of Fan et al [2] and Octree Generating Networks (OGNs) of Tatarchenko et al [3]
We design a new generation network, RealPoint3D, that is more suitable for 3D fine-grained reconstruction from a single image in a real scenario
Summary
Generating 3D point clouds from a single image, which aims to provide solutions to tasks such as autonomous driving, virtual reality and robotic surgery, is a fundamental, intriguing area in remote sensing and computer vision. With the development of deep learning, many learning-based methods have been proposed for depth estimation and 3D reconstruction. 3D reconstruction from a single image considered as an ill-posed problem has achieved promising results by using the deep neural network [2,3,4]. Different from traditional geometric-based approaches [5,6], learning-based methods, with the superpower representation of the deep neural network, build a complex mapping from image space to 3D object space. Before feeding the 3D data into the neural network, the data are usually transformed into volumetric grids or 2D images [7] rendered from different views. Recently-proposed Octree Generating Networks (OGNs) [3] have achieved impressive performances in 3D object generation. There is an obvious disadvantage to the voxel-based method: how to balance sampling resolution and net efficiency is a difficult problem
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have