Zip-NeRF在抗混叠(包括NeRF从空间坐标到颜色和密度的学习映射的空间混叠,以及沿每条射线在线蒸馏过程中使用的损失函数的z-混叠)方面都取得了很好的效果,并且速度相比前作Mip-NeRF 360 提高了24X

mipNeRF 360+基于网格的模型(如Instant NGP)的技术

  • 错误率下降8~77%,并且比Mip-NeRF360提速了24X
  • 主要贡献:
    • 多采样
    • 预滤波

多采样:train左,test右

Image 1
Image 2
Read more »

Reference

[PDF] NeRD: Neural Reflectance Decomposition From Image Collections
[PDF] SAMURAI: Shape And Material from Unconstrained Real-world Arbitrary Image collections
[PDF] Relighting4D: Neural Relightable Human from Videos
[PDF] Neural 3D Scene Reconstruction with the Manhattan-world Assumption
[PDF] NeROIC: Neural Rendering of Objects from Online Image Collections

对金属反光材质的物体重建效果很好

imgae

提出了一种新的光表示方法,颜色由漫反射和镜面反射两部分组成,通过两个阶段的方法来实现

  • Stage1:使用集成方向编码来近似光积分,使用shadow MLP对直接光和间接光进行model,学习到了表面几何形状
  • Stage2:蒙特卡罗采样固定几何形状,重建更精确的表面BRDF和环境光
    • $\mathbf{c}_{\mathrm{diffuse}}=\frac{1}{N_{d}}\sum_{i}^{N_{d}}(1-m)\mathrm{a}L(\omega_{i}),$
    • $\mathbf{c}_{\mathrm{specular}}=\frac{1}{N_{s}}\sum_{i}^{N_{s}}\frac{FG(\omega_{0}\cdot\mathbf{h})}{(\mathbf{n}\cdot\mathbf{h})(\mathbf{n}\cdot\omega_{\mathbf{0}})}L(\omega_{i}),$
Read more »

Title Tri-MipRF: Tri-Mip Representation for Efficient Anti-Aliasing Neural Radiance Fields
Author Wenbo Hu Yuling Wang
Conf/Jour ICCV
Year 2023
Project Tri-MipRF: Tri-Mip Representation for Efficient Anti-Aliasing Neural Radiance Fields (wbhu.github.io)
Paper Tri-MipRF: Tri-Mip Representation for Efficient Anti-Aliasing Neural Radiance Fields (readpaper.com)

2023.7.26 SOTA
image.png|600

image.png

like TensoRF: Tensorial Radiance Fields (apchenstu.github.io)+ Mip-NeRF + NGP

  • 类似Mip-NeRF中的采样锥形方法,但是Tri-MipRF使用了与圆锥相切的采样球S=(x,r),代替Mip-NeRF中的多元高斯圆锥体

    • 采样球的半径通过像素圆盘半径$\dot r$(由像素大小in world Coordinate),焦距f和$t_i$确定
  • 类似TensoRF中的分解方法,将空间采样球分解到三个平面上,编码类似NGP中的HashGrid 使用2D平面来存取特征值,构建一个base level:$M^{L_{0}}$,通过downscaling来获得其他level的2D grid平面。

    • 通过base level中interest space的AABB求出$\ddot r$,并联合采样球半径r得到采样球在平面投影的level,根据此level和投影到平面上的二维坐标,在相邻两level $\mathcal{M}_{XY}^{\lfloor l\rfloor}$和$\mathcal{M}_{XY}^{\lceil l\rceil}$的2D grid中采用3线性插值得到采样球的特征值,最后三个分解平面的特征值cat起来作为MLP的一个输入
  • 一种更好的渲染视图方法:Hybrid Volume-Surface Rendering

    • 通过在密度场中marching cubes和网格抽取来获得代理网格,粗略确定相机原点到物体表面的距离
    • 对代理网格进行有效栅格化,以获得圆锥体中轴线表面上的命中点,然后我们在距圆锥体中轴线命中点∆t距离内均匀采样球体,这产生2∆t采样间隔。
    • 优点:可以减少需要采样点的数量,且不会影响渲染出来图片的质量
  • 优点:

    • fine-grained details in close-up views
    • and free of aliasing in distant views
    • 5 minute and smaller model parameters
  • 缺点:

    • 需要使用multi-view segmentation methods将In-the-wildIn数据集中感兴趣的物体提取出来
      • 即需要mask
Read more »

Title Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields
Author Jonathan T. Barron Ben Mildenhall Dor Verbin Pratul P. Srinivasan Peter Hedman
Conf/Jour CVPR 2022 (Oral Presentation)
Year 2022
Project mip-NeRF 360 (jonbarron.info)
Paper Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields (readpaper.com)
  • novel Kalman-like scene parameterization:将场景参数化,将单位球外背景采样的截头锥参数化到r=2的球体内,单位球内的截头锥不受影响

image.png

  • efficient proposal-based coarse-to-fine distillation framework:一个提议网络用来获取权重,用来进行精采样,再通过精采样的点根据NeRF 的MLP得到密度和颜色值

image.png

  • regularizer designed for mipNeRF ray intervals:可以有效消除floaters(体积密集空间中不相连的小区域)和背景塌陷(远处的表面被错误地模拟成靠近相机的密集内容的半透明云)

$\begin{gathered}\mathcal{L}_{\mathrm{dist}}(\mathbf{s},\mathbf{w}) =\sum_{i,j}w_iw_j\left|\frac{s_i+s_{i+1}}{2}-\frac{s_j+s_{j+1}}{2}\right| \+\frac13\sum_iw_i^2(s_{i+1}-s_i) \end{gathered}$

image.png

Read more »

Title Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields
Author Jonathan T. Barron and Ben Mildenhall and Matthew Tancik and Peter Hedman and Ricardo Martin-Brualla and Pratul P. Srinivasan
Conf/Jour 2021 IEEE/CVF International Conference on Computer Vision (ICCV)
Year 2021
Project google/mipnerf (github.com)
Paper Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields (readpaper.com)

image.png

  • 一种新的采样方式:锥体采样conical frustums截头圆锥体
  • 基于PE提出了IPE,可以平滑地编码空间体积的大小和形状
  • 将NeRF的粗精采样MLP合并为一个MLP

IPE:当锥体区域较宽(正态分布很宽)时,会将高频的信息积分为0;当区域较窄时,保持原来的PEncoding
ipe_anim_horiz.gif

image.png

Read more »

Title NeRF++: Analyzing and Improving Neural Radiance Fields
Author Kai Zhang, Gernot Riegler, Noah Snavely, Vladlen Koltun
Conf/Jour arXiv: Computer Vision and Pattern Recognition
Year 2020
Project Kai-46/nerfplusplus: improves over nerf in 360 capture of unbounded scenes (github.com)
Paper NeRF++: Analyzing and Improving Neural Radiance Fields. (readpaper.com)
arxiv.org/pdf/2010.07492

image.png

创新:一种前背景分离的方法
挑战:

  • First, the training and testing of NeRF and NeRF++ on a single large-scale scene is quite time-consuming and memory-intensive —> NGP解决了耗时
  • Second, small camera calibration errors may impede阻碍 photorealistic synthesis. Robust loss functions, such as the contextual loss (Mechrez et al., 2018), could be applied.
  • Third, photometric effects such as auto-exposure and vignetting渐晕 can also be taken into account to increase image fidelity. This line of investigation is related to the lighting changes addressed in the orthogonal work of Martin-Brualla et al. (2020).
Read more »

Title NeuDA: Neural Deformable Anchor for High-Fidelity Implicit Surface Reconstruction
Author Bowen Cai ,Jinchi Huang, Rongfei Jia ,Chengfei Lv, Huan Fu*
Conf/Jour CVPR
Year 2023
Project NeuDA (3d-front-future.github.io)
Paper NeuDA: Neural Deformable Anchor for High-Fidelity Implicit Surface Reconstruction (readpaper.com)

NeuDA变形后的grid距离Surface更近一些,即可以使采样点插值时更多依赖于表面,即渲染时也会更多地考虑到3D空间相邻的信息

创新:Deformable Anchors、HPE、$\mathcal{L}_{norm}$

  • 改进了NGP中的grid表示,8个顶点存储feature—>存储锚点位置,锚点位置经过PE后输入进SDF网络

image.png

image.png

Read more »

从仅2~3张的稀疏输入中重建表面

  • 首先,我们提出了一个多层几何推理框架,以粗到细的方式恢复表面。
  • 其次,我们采用了一种多尺度颜色混合方案,该方案联合评估局部和背景亮度一致性,以获得更可靠的颜色预测。
  • 第三,采用一致性感知的微调方案,控制遮挡和图像噪声引起的不一致区域,得到准确、干净的重建。

image.png

Read more »

Title PermutoSDF: Fast Multi-View Reconstruction with Implicit Surfaces using Permutohedral Lattices
Author Radu Alexandru Rosu, Sven Behnke
Conf/Jour CVPR
Year 2023
Project PermutoSDF (radualexandru.github.io)
Paper PermutoSDF: Fast Multi-View Reconstruction with Implicit Surfaces using Permutohedral Lattices (readpaper.com)

image.png

创新:用permutohedral lattice替换voxel hash encoding

image.png

Read more »

Title Neuralangelo: High-Fidelity Neural Surface Reconstruction
Author Zhaoshuo LiThomas MüllerAlex EvansRussell H. TaylorMathias UnberathMing-Yu LiuChen-Hsuan Lin
Conf/Jour IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
Year 2023
Project Neuralangelo: High-Fidelity Neural Surface Reconstruction (nvidia.com)
Paper Neuralangelo: High-Fidelity Neural Surface Reconstruction (readpaper.com)

创新:新的计算梯度的方法——数值梯度、粗到精地逐步优化——数值梯度的补偿$\epsilon$,粗网格先激活,当$\epsilon$减小到精网格的空间大小时,逐步激活精网格
SR Issue: Current methods struggle to recover detailed structures of real-world scenes
To address : present Neuralangelo (combines the representation power of multi-resolution 3D hash grids with neural surface rendering)

  • numerical gradients for computing higher-order derivatives as a smoothing operation
    • image.png
  • coarse-to-fine optimization on the hash grids controlling different levels of details
    even wo auxiliary inputs such as depth , Neuralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity 保真 significantly surpassing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures.

our future work to explore a more efficient sampling strategy to accelerate the training process.

Read more »

Title NerfAcc: Efficient Sampling Accelerates NeRFs
Author Li, Ruilong and Gao, Hang and Tancik, Matthew and Kanazawa, Angjoo
Conf/Jour arXiv preprint arXiv:2305.04966
Year 2023
Project NerfAcc Documentation — nerfacc 0.5.3 documentation
Paper NerfAcc: Efficient Sampling Accelerates NeRFs (readpaper.com)

一种可以加速NeRF的高效采样策略
image.png

NerfAcc = Instant-NGP的Occupancy Grid + Mip-NeRF 360的Proposal Network

pip install nerfacc

Read more »

本项目yq010105/NeRF-Mine (github.com)基于Instant-nsr-pl(NSR,NGP,PytorchLightning)代码构建

  • 保留 omegaconf、nerfacc、Mip-nerf_loss,类似文件结构
  • 去除 pytorch-lightning 框架,使用 pytorch

NeRF 主要部分:

  • 神经网络结构 —> 训练出来模型,即 3D 模型的隐式表达
    • 网络类型一般为 MLP,相当于训练一个函数,输入采样点的位置,可以输出该点的信息(eg: density, sdf, color…)
  • 采样方式:沿着光线进行采样获取采样点
  • 位置编码:对采样点的位置 xyz 和方向 dir 进行编码,使得 MLP 的输入为高频的信息
  • 数学相关:光线的生成、坐标变换、体渲染公式、BRDF……
  • 体渲染函数:
    • NeRF:$\mathrm{C}(r)=\int_{\mathrm{t}_{\mathrm{n}}}^{\mathrm{t}_{\mathrm{f}}} \mathrm{T}(\mathrm{t}) \sigma(\mathrm{r}(\mathrm{t})) \mathrm{c}(\mathrm{r}(\mathrm{t}), \mathrm{d}) \mathrm{dt} =\sum_{i=1}^{N} T_{i}\left(1-\exp \left(-\sigma_{i} \delta_{i}\right)\right) \mathbf{c}_{i}$
      • 不透明度$\sigma$,累计透光率 —> 权重
      • 颜色值
    • Neus:$C(\mathbf{o},\mathbf{v})=\int_{0}^{+\infty}w(t)c(\mathbf{p}(t),\mathbf{v})\mathrm{d}t$
      • sdf, dirs, gradients, invs —> $\alpha$ —> 权重
      • 颜色值
    • NeRO:$\mathbf{c}(\omega_{0})=\mathbf{c}_{\mathrm{diffuse}}+\mathbf{c}_{\mathrm{specular}} =\int_{\Omega}(1-m)\frac{\mathbf{a}}{\pi}L(\omega_{i})(\omega_{i}\cdot\mathbf{n})d\omega_{i} + \int_{\Omega}\frac{DFG}{4(\omega_{i}\cdot\mathbf{n})(\omega_{0}\cdot\mathbf{n})}L(\omega_{i})(\omega_{i}\cdot\mathbf{n})d\omega_{i}$
      • 漫反射颜色:Light(直射光),金属度 m、反照率 a
      • 镜面反射颜色:Light(直射光+间接光),金属度 m、反照率 a、粗糙度$\rho$ ,碰撞概率 occ_prob,间接光碰撞 human 的 human_light
      • 详情见NeRO Code
  • 隐式模型导出(.stl、.obj、.ply 等)显式模型(Marching Cube):利用 trimesh,torchmcubes,mcubes 等库
    • 根据 sdf 和 threshold,获取物体表面的 vertices 和 faces(如需还要生成 vertices 对应的 colors)。
    • 然后根据 vertices、faces 和 colors,由 trimesh 生成 mesh 并导出模型为 obj 等格式

Future:

  • [ ] 消除颜色 or 纹理与几何的歧义,Neus(X—>MLP—>SDF)的方法会将物体的纹理建模到物体的几何中
  • [x] 只关注前景物体的建模,可以结合 SAM 将图片中的 interest object 分割出来: Rembg分割后效果也不好
Read more »

Title Human Performance Modeling and Rendering via Neural Animated Mesh
Author Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang Minye Wu, Lan Xu, Jingyi Yu
Conf/Jour SIGGRAPH Asia 2022
Year 2022
Project Human Performance Modeling and Rendering via Neural Animated Mesh (zhaofuq.github.io)
Paper Human Performance Modeling and Rendering via Neural Animated Mesh (readpaper.com)

可以理解为对Neus使用多分辨率哈希编码进行加速

  • 使用TSDF代替SDF
  • 有限差分函数计算SDF的梯度,在tiny-CUDAnn并未集成,公开了自己的CUDAC++代码

pipeline

不足:

  • 数据集需要手动mask
Read more »