面向空间场景建模的三角网格神经辐射场模型压缩  

Triangle Grid Neural Radiation Field Model Compression for Space Scene Modeling

作　　者：李中衡 周轶丁 冯新健 张秋光 王晶 郭宇 王海岳 LI Zhongheng;ZHOU Yiding;FENG Xinjian;ZHANG Qiuguang;WANG Jing;GUO Yu;WANG Haiyue(Beijing Space Vehicle General Design Department,Beijing 100094,China;Beijing Key Laboratory of Space Intelligent Robot System Technology and Application,Beijing 100094,China;National Key Laboratory of Human-machine Hybrid Augmented Intelligence,Xi’an 710048,China;National Engineering Research Center for Visual Information and Applications,Xi’an 710048,China;Institute of Artificial Intelligence and Robotics,Xi’an Jiaotong University,Xi’an 710048,China;School of Software Engineering,Xi’an Jiaotong University,Xi’an 710048,China;School of Energy and Power Engineering,Xi’an Jiaotong University,Xi’an 710048,China)

机构地区：[1]北京空间飞行器总体设计部,北京100094 [2]空间智能机器人系统技术与应用北京市重点实验室,北京100094 [3]人机混合增强智能全国重点实验室,西安710048 [4]视觉信息与应用国家工程研究中心,西安710048 [5]西安交通大学人工智能与机器人研究所,西安710048 [6]西安交通大学软件学院,西安710048 [7]西安交通大学能源与动力工程学院,西安710048

出　　处：《载人航天》2024年第6期844-853,共10页Manned Spaceflight

基　　金：国家自然基金重点项目(61231018)。

摘　　要：为解决神经辐射场渲染方法模型内存占用过大的问题,提出了神经三角网格算法。通过在训练好的三维网格模型中构建顶点特征向量,利用传统渲染管线中的光栅化计算射线与三角网格的交点,并通过重心坐标获得对应交点的特征。在多个公开数据集上的实验结果表明:所提方法在保证渲染质量的前提下,只需要较小的存储空间来存储神经辐射场模型,模型的压缩率达到了25%。可视化的实验结果表明:所提方法能够得到高质量的新视角合成结果。该方法能够剔除场景中无用点的特征和场景内部的特征,减小模型的存储空间,此外三角网格的顶点特征能够直接利用顶点索引获取,从而能够在嵌入式设备上应用。To address the issue of excessive memory usage in neural radiance field rendering meth-ods,a neural triangular mesh algorithm has been proposed.This algorithm constructs vertex feature vectors within trained 3D mesh models and utilizes rasterization in the traditional rendering pipeline to calculate the intersection of rays with the triangular mesh.It obtains the characteristicsof the corresponding intersection points through barycentric coordinates.Experimental results on multiple public datasets demonstrate that this method requires only a small amount of storage space to store the neural radiance field model while ensuring rendering quality,achieving a model compression rate of 25%.Visualized experimental results indicate that this method can produce high-quality novel view synthesis results.The approach can eliminate the features of useless points in the scene and within the scene itself,reducing the model,s storage space.Moreover,the vertex features of the triangular mesh can be directly accessed using vertex indices,making it applicable on embedded devices.

关 键 词：模型压缩 神经辐射场 光栅渲染 神经三角网格 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]