目录

目录

Neural Sparse Voxel Fields


Neural sparse voxel fields

编者按

  • 兼顾精度、速度?
  • 既然是利用形状/点云信息,加快query时的速度,那可否利用SDF?
    • 这种加速,相当于额外存储一下空间中哪些位置占用低

Motivation

  • 现有的方法,由于网络容量/寻找和scene geometry的精确intersection有困难,目前的很多方法都是blurry results
  • 提出一种新的表征,用octree结构的voxels来each 存local feature
    这样在采样时就可以跳过无关的voxels,比NeRF快10倍
    https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20201221173251904.png
  • 至于每个pixel的feature,渲染结构与NeRF类似
    https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20201221174122461.png

Overview

  • self-pruning
    把占用小于阈值的网格直接剪枝掉
    https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20201221174308274.png
  • 一个voxel上的feature有顶点补间得来
    https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20210106152320554.png

Results

  • https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20201221173910965.png
  • 因为用的是显式的sparse voxel 表征,可以轻松用于scene composition
    • https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20201221174604849.png
  • 还可以在 ScanNet 上直接进行尝试:首先用注册后的深度图提取点云
    https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20210106151555886.png
  • 还可以做scene editing:因为用的是显式的voxel
    • https://longtimenohack.com/posts/paper_reading/2020nips_liu_neural/image-20210106151644456.png