Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields
传统神经辐射场(Neural Radiance Field,NeRF)对图像中的每一个像素发射一条光线,在该光线上取若干采样点对场景进行采样。它的弱点在于训练和推导应当尽可能在同一分辨率下,否则会产生模糊(近距离)、混叠(远距离),造成准确率下降。为了解决这个问题,一个直接的想法是增加每个像素发射的光线数量,然而每条光线的渲染都需要查询 MLP 数百次,这种超采样方式的开销是不可接受的。
Mip-NeRF 的做法是用视锥取代光线,借用 CG 中 Mipmap 的概念,利用一些预处理得到的 Mipmap,从而能够更好地还原不同分辨率下的场景。这些 Mipmap 也是由 MLP 学习得到的特征,由于暗含了分辨率参数(事实上,即
t
t
t),Mip-NeRF 只需要一个统一的 MLP 即可表达场景,而不用像传统 NeRF 那样分粗、细粒度两个。
区别于 NeRF 采样空间中无限小的点,Mip-NeRF 采样的是具有形状和大小、对尺度敏感的视锥。对位置向量
x
\bold x
x 参数化:
术语 | 含义 |
---|---|
图像平面( t = 1 t=1 t=1) | o + d {\bold o}+{\bold d} o+d |
图像平面视锥半径 (像素宽度的 2 / 12 2/\sqrt{12} 2/12 倍) | r ˙ {\dot r} r˙ |
视锥范围(近点、远点) | [ t 0 , t 1 ] [t_0,t_1] [t0,t1] |
F
(
x
,
o
,
d
,
r
˙
,
t
0
,
t
1
)
=
1
{
(
t
0
<
d
T
(
x
−
o
)
∥
d
∥
2
2
<
t
1
)
∧
(
d
T
(
x
−
o
)
∥
d
∥
2
∥
x
−
o
∥
2
>
1
1
+
(
r
˙
/
∥
d
∥
2
)
2
)
}
{\rm F}({\bold x},{\bold o},{\bold d},{\dot r},t_0,t_1)=1\left\{\left(t_0<\frac{{\bold d}^{\rm T}({\bold x}-{\bold o})}{\|{\bold d}\|_2^2}<t_1\right)\land\left(\frac{{\bold d}^{\rm T}({\bold x}-{\bold o})}{\|{\bold d}\|_2\|{\bold x}-{\bold o}\|_2}>\frac{1}{\sqrt{1+\left({\dot r}/{\|{\bold d}\|_2}\right)^2}}\right)\right\}
F(x,o,d,r˙,t0,t1)=1⎩
⎨
⎧(t0<∥d∥22dT(x−o)<t1)∧
∥d∥2∥x−o∥2dT(x−o)>1+(r˙/∥d∥2)21
⎭
⎬
⎫用
tan
θ
˙
=
r
˙
∥
d
∥
2
\tan{\dot\theta}=\dfrac{\dot r}{\|{\bold d}\|_2}
tanθ˙=∥d∥2r˙ 表示视锥在图像平面的半张角,上式改写为:
F
(
x
,
o
,
d
,
r
˙
,
t
0
,
t
1
)
=
1
{
(
t
0
<
d
T
(
x
−
o
)
∥
d
∥
2
2
<
t
1
)
∧
(
d
T
(
x
−
o
)
∥
d
∥
2
∥
x
−
o
∥
2
>
cos
θ
˙
)
}
{\rm F}({\bold x},{\bold o},{\bold d},{\dot r},t_0,t_1)=1\left\{\left(t_0<\frac{{\bold d}^{\rm T}({\bold x}-{\bold o})}{\|{\bold d}\|_2^2}<t_1\right)\land\left(\frac{{\bold d}^{\rm T}({\bold x}-{\bold o})}{\|{\bold d}\|_2\|{\bold x}-{\bold o}\|_2}>\cos{\dot\theta}\right)\right\}
F(x,o,d,r˙,t0,t1)=1{(t0<∥d∥22dT(x−o)<t1)∧(∥d∥2∥x−o∥2dT(x−o)>cosθ˙)}当且仅当
F
(
x
,
⋅
)
=
1
{\rm F}({\bold x},\cdot)=1
F(x,⋅)=1 时
x
\bold x
x 在视锥内部。此时位置编码更正为积分位置编码(Integral Positional Encoding,IPE)的期望:
γ
∗
(
o
,
d
,
r
˙
,
t
0
,
t
1
)
=
∫
γ
(
x
)
F
(
x
,
o
,
d
,
r
˙
,
t
0
,
t
1
)
d
x
∫
F
(
x
,
o
,
d
,
r
˙
,
t
0
,
t
1
)
d
x
\gamma^*({\bold o},{\bold d},{\dot r},t_0,t_1)=\frac{\displaystyle\int{\gamma({\bold x}){\rm F}({\bold x},{\bold o},{\bold d},{\dot r},t_0,t_1){\rm d}{\bold x}}}{\displaystyle\int{{\rm F}({\bold x},{\bold o},{\bold d},{\dot r},t_0,t_1){\rm d}{\bold x}}}
γ∗(o,d,r˙,t0,t1)=∫F(x,o,d,r˙,t0,t1)dx∫γ(x)F(x,o,d,r˙,t0,t1)dx上述积分的分子部分没有精确解,Mip-NeRF 使用多元高斯近似(假设视锥沿射线对称)。
术语 | 含义 |
---|---|
射线方向距离均值 | μ t \mu_t μt |
射线方向距离方差 | σ t 2 \sigma_t^2 σt2 |
垂轴方向距离方差 | σ r 2 \sigma_r^2 σr2 |
对位置向量进行坐标变换:
{
x
=
r
′
cos
θ
y
=
r
′
sin
θ
z
=
t
,
1
r
˙
=
t
r
′
⟹
r
′
=
r
˙
t
\begin{cases} x=r'\cos\theta\\ y=r'\sin\theta\\ z=t \end{cases},\quad \frac{1}{{\dot r}}=\frac{t}{r'}\Longrightarrow r'={\dot r}t
⎩
⎨
⎧x=r′cosθy=r′sinθz=t,r˙1=r′t⟹r′=r˙t
x
=
φ
(
r
,
t
,
θ
)
=
(
r
t
cos
θ
,
r
t
sin
θ
,
t
)
,
θ
∈
[
0
,
2
π
)
,
t
≥
0
,
∣
r
∣
≤
r
˙
{\bold x}=\varphi(r,t,\theta)=(rt\cos\theta,rt\sin\theta,t),\theta\in[0,2\pi),t\ge0,|r|\le{\dot r}
x=φ(r,t,θ)=(rtcosθ,rtsinθ,t),θ∈[0,2π),t≥0,∣r∣≤r˙微元:
d
x
d
y
d
z
=
∣
det
D
φ
(
r
,
t
,
θ
)
∣
d
r
d
t
d
θ
=
∣
∂
x
∂
r
∂
x
∂
t
∂
x
∂
θ
∂
y
∂
r
∂
y
∂
t
∂
y
∂
θ
∂
z
∂
r
∂
z
∂
t
∂
z
∂
θ
∣
d
r
d
t
d
θ
=
∣
t
cos
θ
r
cos
θ
−
r
t
sin
θ
t
sin
θ
r
sin
θ
r
t
cos
θ
0
1
0
∣
d
r
d
t
d
θ
=
−
(
r
t
2
cos
2
θ
+
r
t
2
sin
2
θ
)
d
r
d
t
d
θ
=
−
r
t
2
d
r
d
t
d
θ
\begin{align} {\rm d}x{\rm d}y{\rm d}z &=|\det{D\varphi}(r,t,\theta)|{\rm d}r{\rm d}t{\rm d}\theta\\ &=\begin{vmatrix} \dfrac{\partial x}{\partial r} & \dfrac{\partial x}{\partial t} & \dfrac{\partial x}{\partial\theta}\\ \dfrac{\partial y}{\partial r} & \dfrac{\partial y}{\partial t} & \dfrac{\partial y}{\partial\theta}\\ \dfrac{\partial z}{\partial r} & \dfrac{\partial z}{\partial t} & \dfrac{\partial z}{\partial\theta} \end{vmatrix}{\rm d}r{\rm d}t{\rm d}\theta\\ &=\begin{vmatrix} t\cos\theta & r\cos\theta & -rt\sin\theta\\ t\sin\theta & r\sin\theta & rt\cos\theta\\ 0 & 1 & 0 \end{vmatrix}{\rm d}r{\rm d}t{\rm d}\theta\\ &=-(rt^2\cos^2\theta+rt^2\sin^2\theta){\rm d}r{\rm d}t{\rm d}\theta\\ &=-rt^2{\rm d}r{\rm d}t{\rm d}\theta \end{align}
dxdydz=∣detDφ(r,t,θ)∣drdtdθ=
∂r∂x∂r∂y∂r∂z∂t∂x∂t∂y∂t∂z∂θ∂x∂θ∂y∂θ∂z
drdtdθ=
tcosθtsinθ0rcosθrsinθ1−rtsinθrtcosθ0
drdtdθ=−(rt2cos2θ+rt2sin2θ)drdtdθ=−rt2drdtdθ体积:
V
=
∭
0
≤
θ
<
2
π
t
≥
0
∣
r
∣
≤
r
˙
∣
d
x
d
y
d
z
∣
=
∫
0
2
π
d
θ
∫
t
0
t
1
t
2
d
t
∫
0
r
˙
r
d
r
=
2
π
⋅
t
1
3
−
t
0
3
3
⋅
r
˙
2
2
=
π
3
(
t
1
3
−
t
0
3
)
r
˙
2
\begin{align}V&=\iiint\limits_{\begin{aligned}0\le&\theta<2\pi\\&t\ge0\\|&r|\le{\dot r}\end{aligned}}{|{\rm d}x{\rm d}y{\rm d}z|}\\ &=\int_0^{2\pi}{{\rm d}\theta}\int_{t_0}^{t_1}{t^2{\rm d}t\int_0^{\dot r}{r{\rm d}r}}\\ &=2\pi\cdot\frac{t_1^3-t_0^3}{3}\cdot\frac{{\dot r}^2}{2}\\ &=\frac{\pi}{3}(t_1^3-t_0^3){\dot r}^2\end{align}
V=0≤∣θ<2πt≥0r∣≤r˙∭∣dxdydz∣=∫02πdθ∫t0t1t2dt∫0r˙rdr=2π⋅3t13−t03⋅2r˙2=3π(t13−t03)r˙2类似地:
μ
t
=
E
[
t
]
=
1
V
∫
0
2
π
d
θ
∫
t
0
t
1
t
3
d
t
∫
0
r
˙
r
d
r
=
1
V
⋅
2
π
⋅
t
1
4
−
t
0
4
4
⋅
r
˙
2
2
=
3
(
t
1
4
−
t
0
4
)
4
(
t
1
3
−
t
0
3
)
\mu_t={\rm E}[t]=\frac1V\int_0^{2\pi}{{\rm d}\theta}\int_{t_0}^{t_1}{t^3{\rm d}t\int_0^{\dot r}{r{\rm d}r}}=\frac1V\cdot2\pi\cdot\frac{t_1^4-t_0^4}{4}\cdot\frac{{\dot r}^2}{2}=\frac{3(t_1^4-t_0^4)}{4(t_1^3-t_0^3)}
μt=E[t]=V1∫02πdθ∫t0t1t3dt∫0r˙rdr=V1⋅2π⋅4t14−t04⋅2r˙2=4(t13−t03)3(t14−t04)
E
[
t
2
]
=
1
V
∫
0
2
π
d
θ
∫
t
0
t
1
t
4
d
t
∫
0
r
˙
r
d
r
=
1
V
⋅
2
π
⋅
t
1
5
−
t
0
5
5
⋅
r
˙
2
2
=
3
(
t
1
5
−
t
0
5
)
5
(
t
1
3
−
t
0
3
)
{\rm E}[t^2]=\frac1V\int_0^{2\pi}{{\rm d}\theta}\int_{t_0}^{t_1}{t^4{\rm d}t\int_0^{\dot r}{r{\rm d}r}}=\frac1V\cdot2\pi\cdot\frac{t_1^5-t_0^5}{5}\cdot\frac{{\dot r}^2}{2}=\frac{3(t_1^5-t_0^5)}{5(t_1^3-t_0^3)}
E[t2]=V1∫02πdθ∫t0t1t4dt∫0r˙rdr=V1⋅2π⋅5t15−t05⋅2r˙2=5(t13−t03)3(t15−t05)
σ
t
2
=
V
a
r
[
t
]
=
E
[
t
2
]
−
E
2
[
t
]
=
3
(
t
1
5
−
t
0
5
)
5
(
t
1
3
−
t
0
3
)
−
μ
t
2
\sigma_t^2={\rm Var}[t]={\rm E}[t^2]-{\rm E}^2[t]=\frac{3(t_1^5-t_0^5)}{5(t_1^3-t_0^3)}-\mu_t^2
σt2=Var[t]=E[t2]−E2[t]=5(t13−t03)3(t15−t05)−μt2
E
[
x
]
=
1
V
∫
0
2
π
cos
θ
d
θ
∫
t
0
t
1
t
3
d
t
∫
0
r
˙
r
2
d
r
=
0
{\rm E}[x]=\frac1V\int_0^{2\pi}{\cos\theta{\rm d}\theta}\int_{t_0}^{t_1}{t^3{\rm d}t\int_0^{\dot r}{r^2{\rm d}r}}=0
E[x]=V1∫02πcosθdθ∫t0t1t3dt∫0r˙r2dr=0
E
[
x
2
]
=
1
V
∫
0
2
π
cos
2
θ
d
θ
∫
t
0
t
1
t
4
d
t
∫
0
r
˙
r
3
d
r
=
1
V
⋅
π
⋅
t
1
5
−
t
0
5
5
⋅
r
˙
4
4
=
3
(
t
1
5
−
t
0
5
)
20
(
t
1
3
−
t
0
3
)
r
˙
2
{\rm E}[x^2]=\frac1V\int_0^{2\pi}{\cos^2\theta{\rm d}\theta}\int_{t_0}^{t_1}{t^4{\rm d}t\int_0^{\dot r}{r^3{\rm d}r}}=\frac1V\cdot\pi\cdot\frac{t_1^5-t_0^5}{5}\cdot\frac{{\dot r}^4}{4}=\frac{3(t_1^5-t_0^5)}{20(t_1^3-t_0^3)}{\dot r}^2
E[x2]=V1∫02πcos2θdθ∫t0t1t4dt∫0r˙r3dr=V1⋅π⋅5t15−t05⋅4r˙4=20(t13−t03)3(t15−t05)r˙2
σ
r
2
=
V
a
r
[
x
]
=
E
[
x
2
]
−
E
2
[
x
]
=
3
(
t
1
5
−
t
0
5
)
20
(
t
1
3
−
t
0
3
)
r
˙
2
\sigma_r^2={\rm Var}[x]={\rm E}[x^2]-{\rm E}^2[x]=\frac{3(t_1^5-t_0^5)}{20(t_1^3-t_0^3)}{\dot r}^2
σr2=Var[x]=E[x2]−E2[x]=20(t13−t03)3(t15−t05)r˙2上述公式在数值计算层面存在指数爆炸(五次方)、差为零(
t
1
−
t
0
t_1-t_0
t1−t0 太小)的问题,训练过程会产生 0 或者
N
a
N
\rm NaN
NaN,需要进一步参数化
t
1
t_1
t1 和
t
0
t_0
t0 :
{
t
μ
=
t
1
+
t
0
2
t
δ
=
t
1
−
t
0
2
\begin{cases} t_\mu=\dfrac{t_1+t_0}{2}\\ t_\delta=\dfrac{t_1-t_0}{2} \end{cases}
⎩
⎨
⎧tμ=2t1+t0tδ=2t1−t0重写前述公式:
μ
t
=
3
4
⋅
(
t
1
−
t
0
)
(
t
1
+
t
0
)
[
(
t
1
+
t
0
)
2
−
2
t
1
t
0
]
(
t
1
−
t
0
)
(
t
1
2
+
t
0
2
+
t
1
t
0
)
=
3
(
t
μ
2
+
t
δ
2
)
t
μ
3
t
μ
2
+
t
δ
2
=
t
μ
+
2
t
μ
t
δ
2
3
t
μ
2
+
t
δ
2
\begin{align}\mu_t&=\dfrac34\cdot\dfrac{(t_1-t_0)(t_1+t_0)\left[(t_1+t_0)^2-2t_1t_0\right]}{(t_1-t_0)(t_1^2+t_0^2+t_1t_0)}\\ &=\dfrac{3(t_\mu^2+t_\delta^2)t_\mu}{3t_\mu^2+t_\delta^2}\\ &=t_\mu+\dfrac{2t_\mu t_\delta^2}{3t_\mu^2+t_\delta^2} \end{align}
μt=43⋅(t1−t0)(t12+t02+t1t0)(t1−t0)(t1+t0)[(t1+t0)2−2t1t0]=3tμ2+tδ23(tμ2+tδ2)tμ=tμ+3tμ2+tδ22tμtδ2
σ
t
2
=
3
5
⋅
(
t
1
−
t
0
)
[
(
t
1
2
+
t
0
2
)
2
−
(
t
1
t
0
)
2
+
(
t
1
2
+
t
0
2
)
t
1
t
0
]
(
t
1
−
t
0
)
(
t
1
2
+
t
0
2
+
t
1
t
0
)
−
μ
t
2
=
3
(
5
t
μ
4
−
2
t
μ
2
t
δ
2
+
t
δ
4
)
t
δ
2
5
(
3
t
μ
2
+
t
δ
2
)
2
=
[
1
3
−
4
(
12
t
μ
2
−
t
δ
2
)
t
δ
2
15
(
3
t
μ
2
+
t
δ
2
)
2
]
t
δ
2
\begin{align} \sigma_t^2&=\dfrac35\cdot\dfrac{(t_1-t_0)\left[\left(t_1^2+t_0^2\right)^2-\left(t_1 t_0\right)^2+\left(t_1^2+t_0^2\right)t_1t_0\right]}{(t_1-t_0)(t_1^2+t_0^2+t_1t_0)}-\mu_t^2\\ &=\dfrac{3(5t_\mu^4-2t_\mu^2t_\delta^2+t_\delta^4)t_\delta^2}{5\left(3t_\mu^2+t_\delta^2\right)^2}\\ &=\left[\dfrac13-\dfrac{4(12t_\mu^2-t_\delta^2)t_\delta^2}{15\left(3t_\mu^2+t_\delta^2\right)^2}\right]t_\delta^2 \end{align}
σt2=53⋅(t1−t0)(t12+t02+t1t0)(t1−t0)[(t12+t02)2−(t1t0)2+(t12+t02)t1t0]−μt2=5(3tμ2+tδ2)23(5tμ4−2tμ2tδ2+tδ4)tδ2=[31−15(3tμ2+tδ2)24(12tμ2−tδ2)tδ2]tδ2
σ
r
2
=
3
20
⋅
(
t
1
−
t
0
)
[
(
t
1
2
+
t
0
2
)
2
−
(
t
1
t
0
)
2
+
(
t
1
2
+
t
0
2
)
t
1
t
0
]
(
t
1
−
t
0
)
(
t
1
2
+
t
0
2
+
t
1
t
0
)
⋅
r
˙
2
=
(
t
μ
2
4
+
5
t
δ
2
12
−
4
t
μ
2
t
δ
2
15
(
3
t
μ
2
+
t
δ
2
)
)
r
˙
2
\begin{align} \sigma_r^2&=\dfrac{3}{20}\cdot\dfrac{(t_1-t_0)\left[\left(t_1^2+t_0^2\right)^2-\left(t_1 t_0\right)^2+\left(t_1^2+t_0^2\right)t_1t_0\right]}{(t_1-t_0)(t_1^2+t_0^2+t_1t_0)}\cdot{\dot r}^2\\ &=\left(\dfrac{t_\mu^2}{4}+\dfrac{5t_\delta^2}{12}-\dfrac{4t_\mu^2t_\delta^2}{15(3t_\mu^2+t_\delta^2)}\right){\dot r}^2 \end{align}
σr2=203⋅(t1−t0)(t12+t02+t1t0)(t1−t0)[(t12+t02)2−(t1t0)2+(t12+t02)t1t0]⋅r˙2=(4tμ2+125tδ2−15(3tμ2+tδ2)4tμ2tδ2)r˙2于是相机坐标到世界坐标的变换:
{
μ
=
o
+
μ
t
d
Σ
=
σ
t
2
d
d
T
+
σ
r
2
(
I
−
d
d
T
∥
d
∥
2
2
)
\begin{cases} {\boldsymbol\mu}={\bold o}+\mu_t{\bold d}\\ {\boldsymbol\Sigma}=\sigma_t^2{\bold d}{\bold d}^{\rm T}+\sigma_r^2\left({\bold I}-\dfrac{{\bold d}{\bold d}^{\rm T}}{\|{\bold d}\|_2^2}\right) \end{cases}
⎩
⎨
⎧μ=o+μtdΣ=σt2ddT+σr2(I−∥d∥22ddT)矩阵变换形式的位置编码:
P
=
[
1
0
0
2
0
0
2
L
−
1
0
0
0
1
0
0
2
0
⋯
0
2
L
−
1
0
0
0
1
0
0
2
0
0
2
L
−
1
]
T
,
γ
(
x
)
=
[
sin
P
x
cos
P
x
]
{\bold P}=\begin{bmatrix} 1 & 0 & 0 & 2 & 0 & 0 & & 2^{L-1} & 0 & 0 \\ 0 & 1 & 0 & 0 & 2 & 0 & \cdots & 0 & 2^{L-1} & 0 \\ 0 & 0 & 1 & 0 & 0 & 2 & & 0 & 0 & 2^{L-1} \end{bmatrix}^{\rm T}, \gamma({\bold x})=\begin{bmatrix} \sin{{\bold P}{\bold x}}\\ \cos{{\bold P}{\bold x}} \end{bmatrix}
P=
100010001200020002⋯2L−10002L−10002L−1
T,γ(x)=[sinPxcosPx]根据:
E
x
∼
N
(
μ
,
Σ
)
[
sin
x
cos
x
]
=
[
sin
μ
exp
(
−
1
2
σ
2
)
cos
μ
exp
(
−
1
2
σ
2
)
]
{\rm E}_{x\sim{\cal N({\boldsymbol\mu},{\boldsymbol\Sigma})}}\begin{bmatrix} \sin{x}\\ \cos{x} \end{bmatrix}=\begin{bmatrix} \sin{\mu}\exp\left(-\dfrac{1}{2}\sigma^2\right)\\ \cos{\mu}\exp\left(-\dfrac{1}{2}\sigma^2\right) \end{bmatrix}
Ex∼N(μ,Σ)[sinxcosx]=
sinμexp(−21σ2)cosμexp(−21σ2)
定义积分位置编码:
γ
(
μ
′
,
Σ
′
)
=
E
x
∼
N
(
μ
′
,
Σ
′
)
[
γ
(
x
)
]
=
[
sin
μ
′
∘
exp
(
−
1
2
d
i
a
g
Σ
′
)
cos
μ
′
∘
exp
(
−
1
2
d
i
a
g
Σ
′
)
]
\gamma({\boldsymbol\mu}',{\boldsymbol\Sigma}')={\rm E}_{{\bold x}\sim{\cal N({\boldsymbol\mu}',{\boldsymbol\Sigma}')}}[\gamma({\bold x})]=\begin{bmatrix} \sin{{\boldsymbol\mu}'}\circ\exp\left(-\dfrac{1}{2}{\rm diag\ }{\boldsymbol\Sigma}'\right)\\ \cos{{\boldsymbol\mu}'}\circ\exp\left(-\dfrac{1}{2}{\rm diag\ }{\boldsymbol\Sigma}'\right) \end{bmatrix}
γ(μ′,Σ′)=Ex∼N(μ′,Σ′)[γ(x)]=
sinμ′∘exp(−21diag Σ′)cosμ′∘exp(−21diag Σ′)
其中:
{
μ
′
=
P
μ
Σ
′
=
P
Σ
P
T
\begin{cases} {\boldsymbol\mu}'={\bold P}{\boldsymbol\mu}\\ {\boldsymbol\Sigma}'={\bold P}{\boldsymbol\Sigma}{\bold P}^{\rm T} \end{cases}
{μ′=PμΣ′=PΣPT
d
i
a
g
Σ
′
=
[
d
i
a
g
Σ
4
d
i
a
g
Σ
⋯
4
L
−
1
d
i
a
g
Σ
]
T
{\rm diag\ }{\boldsymbol\Sigma}'= \begin{bmatrix} {\rm diag\ }{\boldsymbol\Sigma} & 4{\rm diag\ }{\boldsymbol\Sigma} & \cdots & 4^{L-1}{\rm diag\ }{\boldsymbol\Sigma} \end{bmatrix}^{\rm T}
diag Σ′=[diag Σ4diag Σ⋯4L−1diag Σ]T于是:
γ
(
μ
′
,
Σ
′
)
=
{
[
sin
2
l
μ
∘
exp
(
−
2
2
l
−
1
d
i
a
g
Σ
)
cos
2
l
μ
∘
exp
(
−
2
2
l
−
1
d
i
a
g
Σ
)
]
}
l
=
0
L
\gamma({\boldsymbol\mu}',{\boldsymbol\Sigma}')=\left\{\begin{bmatrix} \sin{2^l{\boldsymbol\mu}}\circ\exp\left(-2^{2l-1}{\rm diag\ }{\boldsymbol\Sigma}\right)\\ \cos{2^l{\boldsymbol\mu}}\circ\exp\left(-2^{2l-1}{\rm diag\ }{\boldsymbol\Sigma}\right) \end{bmatrix}\right\}_{l=0}^L
γ(μ′,Σ′)={[sin2lμ∘exp(−22l−1diag Σ)cos2lμ∘exp(−22l−1diag Σ)]}l=0L由于 IID,我们只需要关心协方差矩阵的对角线,
Σ
\boldsymbol\Sigma
Σ 的计算可以简化为:
d
i
a
g
Σ
=
σ
t
2
(
d
∘
d
)
+
σ
r
2
(
1
−
d
∘
d
∥
d
∥
2
2
)
{\rm diag\ }{\boldsymbol\Sigma}=\sigma_t^2\left({\bold d}\circ{\bold d}\right)+\sigma_r^2\left(1-\dfrac{{\bold d}\circ{\bold d}}{\|{\bold d}\|_2^2}\right)
diag Σ=σt2(d∘d)+σr2(1−∥d∥22d∘d)
References
J. T. Barron, B. Mildenhall, M. Tancik, et al. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. ICCV’21. https://arxiv.org/abs/2103.13415