| 摘 要: 针对基于神经辐射场的渲染方法虽然具备低人工参与度下的照片级别图像生成能力,但是生成图像的时间过长、难以实现实时渲染的问题,文章聚焦于提升神经辐射场的实时渲染性能,从神经辐射场体渲染环节处着手,以烘焙数据为渲染资产,针对利用八叉树保存体素数据无法达到常数访问时间复杂度的问题,提出了一种基于八叉树的扁平化稀疏体素存储方式,以及相应的渲染采样算法。实验结果表明,在使用神经辐射场(Neural Radiance Field,NeRF)合成数据集渲染800×800分辨率的图像时,可以达到268.83的平均帧率,高于其他对比方法。 |
| 关键词: 神经辐射场;渲染;稀疏体素;空间数据结构优化 |
|
中图分类号: TP37
文献标识码: A
|
| 基金项目: 浙江省重点研发项目(2022C01079,2024C01060) |
|
| A Method for Accelerating Rendering of NeRF Baked Data |
|
WANG Xiaomeng1,3, FANG Mengyuan2,3
|
(1.School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.School of In f ormation Science and Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China;
3.Key Laboratory o f Intelligent Textile and Flexible Interconnection o f Zhejiang Province, Hangzhou 310018, China)
345611932@qq.com; myfang@zstu.edu.cn
|
| Abstract: Rendering methods based on NeRF(Neural Radiance Field) demonstrates the capability to generate photorealistic images with minimal human involvement. However, the long generation time and difficulty in achieving real-time rendering pose significant challenges. This paper focuses on enhancing the real-time rendering performance of neural radiance fields, targeting the rendering phase of the neural radiance field and utilizing baked data as rendering assets. To address the issue of failing to achieve constant-time access complexity when using octrees to store voxel, a flattened sparse voxel storage method based on octree is proposed, along with a corresponding rendering sampling algorithm. Experimental results show that when the NeRF synthesized dataset is used to render images at a resolution of 800×800, an average frame rate of 268.83 can be achieved, outperforming other comparison methods. |
| Keywords: NeRF; rendering; sparse voxel; spatial data structure optimization |
