首页|一种基于隐式表征的即时实景三维重建与神经渲染方法

一种基于隐式表征的即时实景三维重建与神经渲染方法

扫码查看
针对神经辐射场训练与渲染速度慢、真实场景渲染视图清晰度低与几何重建效果不佳等问题,该文提出一种基于隐式表征的即时实景三维重建与神经渲染方法.该方法通过使用组合式的优化采样结构以及光场网络的嵌入外观编码,提升了真实场景的视图渲染质量;利用多分辨率哈希表对隐式特征网格进行位置编码,减少了编码字典大小和特征级别数量,此外采用优化的截断符号距离算法,表征真实场景的隐式几何信息.在标准数据集以及无人机航摄影像数据集上,对比该文方法与其他方法的重建性能.结果表明,综合模型精度、重建与渲染时间,该文方法优于文中的其他方法,能够快速地构建隐式表征模型,并即时渲染高度细节还原的实景视图以及重建高质量的实景三维模型.
An instant real scene 3D reconstruction and neural rendering method based on implicit representation
To address issues such as slow training and rendering speed of neural radiation fields,poor rendering quality of real scene views,and poor geometric reconstruction effects,an instant real scene 3D reconstruction and neural rendering method based on implicit representation was proposed in this paper.The view rendering quality of real scenes was improved by using a combination of optimized sampling structures and embedded appearance encoding of light field networks;a multi-resolution hash table was used to position encode the implicit feature grid reduces the size of the encoding dictionary and the number of feature levels.In addition,an optimized truncated symbol distance algorithm was used to represent the implicit geometric information of the real scene.The reconstruction performance of the proposed method was compared with other methods on standard datasets and unmanned aerial photography datasets.The results showed that the proposed method outperformed other methods in terms of model accuracy,reconstruction and rendering time,which could quickly construct implicit representation models,instantly render highly detailed restored real-life views,and reconstruct high-quality real-life 3D models.

virtual realityneural radiation fieldimplicit representationrealistic 3Dneural rendering

吴双品、马劲松、佘江峰

展开 >

南京大学地理与海洋科学学院/江苏省地理信息技术重点实验室/自然资源部国土卫星遥感应用重点实验室,南京 210023

虚拟现实 神经辐射场 隐式表征 实景三维 神经渲染

国家自然科学基金项目

41871293

2024

测绘科学
中国测绘科学研究院

测绘科学

CSTPCD北大核心
影响因子:0.774
ISSN:1009-2307
年,卷(期):2024.49(4)
  • 3