Point Cloud Representation Based On Neural Implicit Functions And Application

Point cloud representation is a crucial research area in the 3D vision domain.High-quality point clouds are essential as input for 3D vision applications such as autonomous driving,AR/VR,3D surface reconstruction,and robotic perception.Due to hardware scanning device limitations,like Li DAR sensors and RGBD depth cameras,the raw point cloud data acquired is sparse,unevenly distributed,and noisy,significantly impeding the progress of these 3D vision applications.In recent years,deep learning,with its data-driven nature and potent learning capabilities,has been extensively employed in the point cloud field.To obtain high quality point clouds,researchers use neural networks to upsample raw point cloud coordinates,resulting in denser and more evenly distributed output points,thereby improving point cloud representation capacity to a certain extent.However,the point cloud upsampling task solely concentrates on point cloud geometry and overlooks texture,which is indispensable for 3D vision tasks such as point cloud rendering.Existing point cloud geometry and texture representation methods can only output fixed-resolution point cloud models,with representation capabilities still constrained by resolution.To surmount the resolution limitation for point cloud geometry and texture representation,this paper first designs a point cloud representation based on a neural implicit function.Utilizing neural implicit functions,each point in the point cloud is represented as a continuous local surface patch centered on that point,replete with rich geometric details and high-fidelity texture.This breakthrough transforms the existing "discrete-discrete" point cloud geometry and texture representation method into a "discrete-continuous" point cloud geometry and texture representation,further augmenting the representation capability of point clouds.Furthermore,since no benchmark dataset exists for point cloud geometry and texture representation methods,this paper establishes a large-scale patch-based dataset for point cloud geometry and texture upsampling tasks and offers a detailed description of the dataset creation process.Additionally,an end-to-end deep learning pipeline based on neural implicit functions is designed,and a point cloud representation based on a neural implicit function is applied to point cloud geometry and texture upsampling tasks,enabling arbitrary resolution upsampling for both point cloud geometry and texture.Experimental results demonstrate that the designed algorithm pipeline outperforms existing fixed-resolution model point cloud representation pipelines in both point cloud geometry and texture metrics,effectively capturing the rich geometric details and high-fidelity texture of point clouds.