Slides

advertisement

Light Fields

PROPERTIES AND APPLICATIONS

Outline

 What are light fields

 Acquisition of light fields

 from a 3D scene

 from a real world scene

 Image rendering from light fields

 Changing viewing angle

 Changing the focal plane

 Sampling and reconstruction

 Depth vs spectral support

 Optimal reconstruction

 Analysis of light transport

Outline

 What are light fields

 Acquisition of light fields

 from a 3D scene

 from a real world scene

 Image rendering from light fields

 Changing viewing angle

 Changing the focal plane

 Sampling and reconstruction

 Depth vs spectral support

 Optimal reconstruction

 Analysis of light transport

The Plenoptic Function

Plenus – Complete, full.

Optic - appearance, look.

The set of things one can ever see

𝑷(𝒙, 𝒚, 𝒛, 𝜽, 𝝓, 𝒕, 𝝀)

Light intensity as a function of

◦ Viewpoint – orientation and position

◦ Time

◦ Wavelength

7D function!

The 5D Plenoptic Function

Ignoring wavelength and time

We need a 5D function to describe light rays across occlusions

◦ 2D orientation

◦ 3D position

The Light Field (4D

Assuming no occlusions

◦ Light is constant across rays

◦ Need only 4D to represent the space of Rays

Is this assumption reasonable?

In free space, i.e outside the convex hull of the scene occluders

The Light Field

Parameterizations

◦ Point on a Plane or curved Surface (2D) and Direction on a Hemisphere (2D)

◦ Two Points on a Sphere

◦ Two Points on two different Planes

Two Plane Parameterization

Convenient parameterization for computational photography

Why?

• Similar to camera geometry (i.e. film plane vs lens plane)

• Linear parameterization - easy computations , no trigonometric functions, etc.

2D light field

Used for visualization. Assume the world is flat (2D)

Intuition

The image a pinhole at

(u,v) captures

𝐼(𝑢, 𝑣) = 𝐿(: , : , 𝑢, 𝑣)

𝐿(𝑠, 𝑡, : , : )

All views of a pixel (s,t)

Light Field Rendering , Levoy Hanrahan '96.

Outline

 What are light fields

 Acquisition of light fields

 from a 3D scene

 from a real world scene

 Image rendering from light fields

 Changing viewing angle

 Changing the focal plane

 Sampling and reconstruction

 Depth vs spectral support

 Optimal reconstruction

 Analysis of light transport

Acquisition of Light Fields

Synthetic 3D Scene

◦ Discretize s,t,u,v and capture all rays intersecting the objects using a standard Ray Tracer

Acquisition of Light Fields

Real world scenes

Will be explained in more detail next week…

Outline

 What are light fields

 Acquisition of light fields

 from a 3D scene

 from a real world scene

 Image rendering from light fields

 Changing viewing angle

 Changing the focal plane

 Sampling and reconstruction

 Depth vs spectral support

 Optimal reconstruction

 Analysis of light transport

Changing the View Point

Problem: Computer Graphics

◦ Render a novel view point without expensive ray tracing

Solution:

◦ Sample a Synthetic light field using Ray Tracing

◦ Use the Light Field to generate any point of view, no need to Ray Trace

Light Field Rendering , Levoy Hanrahan '96.

Changing the View Point

Conceptually: Use Ray Trace from all pixels in image plane pinhole

Actually: Use Homographic mapping from XY plane to the VU and TS, and lookup resulting ray radiance.

Light Field Interpolation

𝐼 𝑋, 𝑌 = 𝐿(𝐻

𝑈𝑉 𝑥, 𝑦 , 𝐻

𝑆𝑇

Problem: Finite sampling of the Light Field –

◦ 𝐻

𝑈𝑉 𝑥, 𝑦 , 𝐻

𝑆𝑇 𝑥, 𝑦 may not be sampled 𝑥, 𝑦 )

Solution: Proper interpolation / reconstruction is needed

◦ Nearest neighbor,

◦ Linear,

◦ Custom Filter

Detailed Analysis later on…

NN NN + Linear Linear

Changing the focal plane

Fourier Slice Photography , Ng, 05

In-Camera Light Field Parameterization

The camera operator

Can define a camera as an operator on the Light Field.

◦ The conventional camera operator:

[Stroebel et al. 1986] x y

Reminder - Thin lens formula

1

+

1

D’ D

= const

D

D’

To focus closer - increase the sensor-to-lens distance .

Refocusing - Reparameterization

Type equation here.

𝐿

𝐹

′ 𝑢, 𝑥 = 𝐿

𝐹

(𝑢, 𝑢 + 𝑥 − 𝑢 𝛼

)

Reparametrization - 4D

Refocusing - Reparameterization

Refocus

Change of distance between planes

𝐹

= 𝛼𝐹

Reparameterization of the light field

Shearing of the Light field

Refocusing camera operator

Shear and Integrate the original light field

*(cos term from conventional camera model is absorbed into L)

Computation of Refocusing Operator

• Naïve Approach

• For every X,Y go over all U,V and calculate the sum after reparameterization => O(n^4) y

′ ′

′ x

• Can we do better ????

Fourier Slice Theorem

𝐹 ∘ 𝐼 = 𝑆 ∘ 𝐹

• F – Fourier Transform Operator

• I – Integral Projection Operator

• S – Slicing Operator

Fourier Analysis of the Camera Operator

Recall that the Refocusing Camera Operator is:

And from the Last theorem we get The Fourier Slice Photography Theorem

Better Algorithm! 𝑃𝑟𝑒𝑝𝑟𝑜𝑐𝑒𝑠𝑠: 𝑂 𝑛 4 log 𝑛 . 𝐶𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒 𝑛𝑒𝑤 𝐷𝑒𝑝𝑡ℎ 𝛼: 𝑂 𝑛 2 + 𝑂(𝑛 2 log 𝑛)

Fourier Slice Photography , Ng, 05

Fourier Slice Photography Thm – More corollaries

Two important results that are worth mentioning:

1. Filtered Light Field Photography Thm

𝑳

𝑭

′ *K=?

2. The light field dimensionality gap

Filtered Light Field Photography Thm

Theorem: Filtered Light Field Photography

The light field dimensionality gap

◦ The light field is 4D

◦ In the frequency domain – The support of all the images with different focus depth is a 3D manifold

This observation was used in order to generate new views of the scene from a focal stack (Levin et al. 2010)

Outline

 What are light fields

 Acquisition of light fields

 from a 3D scene

 from a real world scene

 Image rendering from light fields

 Changing viewing angle

 Changing the focal plane

 Sampling and reconstruction

 Depth vs spectral support

 Optimal reconstruction

 Analysis of light transport

Light Field Sampling

• Light Field Acquisition – Discretization

• Light Field Sampling is Limited

Example – Camera Array: t,s u,v

Sampling in frequency domain

Aliasing in the frequency domain

*

Need to analyze Light Field Spectrum

=

Scene Depth and Light Field

 Light Field Spectrum is related to Scene Depth

 From Lambertian property each point in the scene corresponds to a line in the Light Field

 Line slope is a function of the depth (z) of the point.

Plenoptic Sampling , Chai et al., 00.

Spectral Support of Light Field

Constant Depth

Scene Light Field LF Spectrum

Plenoptic Sampling , Chai et al., 00.

Spectral Support of Light Field

Varying Depth

Scene LF Spectrum

Plenoptic Sampling , Chai et al., 00.

Spectral Support of Light Field

Plenoptic Sampling , Chai et al., 00.

Reconstruction Filters

Optimal Slope for filter:

Plenoptic Sampling , Chai et al., 00.

Limitations

Assumptions

◦ Lambertian surfaces

◦ Free Space – No occlusions

Frequency Analysis of Light Transport

• Informally: Different features of lighting and scene causes different effects in the Frequency

Content

• Blurry Reflections

• Shadow Boundries

High frequency Low frequency

A Frequency analysis of Light Transport , Durand et al. 05.

Not Wave Optics!!!

Frequency Analysis of Light Transport

Look at light transport as a signal processing system.

◦ Light source is the input signal

◦ Interaction are filters / transforms

Source Transport Occlusion Transport Reflection

(BRDF)

Local Light Field

We study the local 4D Light Field around a central Ray during transport

◦ In Spatial Domain

◦ In Frequency Domain

* Local light field offers us the ability to talk about the Spectrum

In a local setting

Local Light Field (2D) Parameterization

The analysis is in flatland, an extension to 4D light field is available x-v parameterization x-Θ parameterization

A Frequency analysis of Light Transport , Durand et al. 05.

Example Scenario

Reflection

A Frequency analysis of Light Transport , Durand et al. 05.

Light Transport – Spatial Domain

Light Propagation  Shear of the local Light Field

◦ No change in slope (v)

◦ Linear change in displacement (X)

+

Light Transport – Frequency Domain

Shear in spatial domain is also a shear in Frequency domain

Occlusion

Spatial domain:

Occlusion  pointwise multiplication in the spatial domain

The incoming light field is multiplied by the binary occlusion function of the occluders.

Frequency domain convolution in the frequency domain:

Occlusion – example

Reflection

We consider planar surfaces * and rotation invariant BRDFs here

What happens when light hits a surface?

1. Multiplication by a cosine term 𝑙 𝑥, 𝜃 = 𝑙 𝑥, 𝜃 cos

+

(𝜃)

2. Mirror Reparameterization around the normal direction 𝑙 𝑥, 𝜃 = 𝑙 𝑥, −𝜃

3. convolution with the BRDF 𝑙 𝑥, 𝜃 = 𝑙 𝑥, 𝜃 ∗ 𝑏𝑟𝑑𝑓(𝜃)

* Similar analysis for curved surfaces is also presented in the paper

Reflection - cosine term

Spatial domain - multiplication:: 𝑙

𝑅′ 𝑥, 𝜃 = 𝑙

𝑅 𝑥, 𝜃 cos

+

(𝜃)

Frequency domain:

𝑙

𝑅′

Ω 𝑥

, Ω 𝜃

Ω 𝑥

, Ω 𝜃

∗ 𝐹(cos

+ 𝜃 )

cosine term example

Incoming

Light field

Light field

After cosine term

Reflection – Mirror reparameterization

Mirror Reparameterization around the normal direction

◦ Using the law of reflection 𝜃 𝑖𝑛

= 𝜃 𝑜𝑢𝑡

◦ 𝑙

𝑅′ 𝑥, 𝜃 = 𝑙

𝑅 𝑥, −𝜃 𝜃 𝑖𝑛 𝜃 𝑜𝑢𝑡

Frequency domain: mirror in the spatial domain => mirror in the frequency domain 𝑙

𝑅′

Ω 𝑥

, Ω 𝜃

= 𝑙

𝑅

Ω 𝑥

, −Ω 𝜃

Reparameterization example

Incoming

Light field

Light field

After reparameterization

𝜋

2

Reflection - BRDF

What is a BRDF?

◦ Bidirectional reflectance distribution function

◦ A function of the incoming ant out going angles 𝜌(𝜃 𝑖𝑛

, 𝜃 𝑜𝑢𝑡

)

◦ Tells us how much “light” comes out at a angle 𝜃 𝑜𝑢𝑡 when illuminating the point from 𝜃 𝑖𝑛

.

◦ Different BRDFs model the reflectance properties of different materials

◦ A lot of BRDFs depend only on the difference between 𝜃 𝑜𝑢𝑡 and the mirror reflection direction: 𝜌 𝜃 𝑖𝑛

, 𝜃 𝑜𝑢𝑡

= 𝑐𝑜𝑛𝑠𝑡 𝜌 𝜃 𝑖𝑛

, 𝜃 𝑜𝑢𝑡

= 𝛿 𝜃 𝑟𝑒𝑓𝑙𝑒𝑐𝑡

− 𝜃 𝑜𝑢𝑡 𝜌 𝜃 𝑖𝑛

, 𝜃 𝑜𝑢𝑡

= 𝑓 𝜃 𝑟𝑒𝑓𝑙𝑒𝑐𝑡

− 𝜃 𝑜𝑢𝑡 𝜋

2 𝜋

2 𝜋

2 𝜋

2 𝜋

2

BRDF Intuition

Assume a Specular BRDF, flat surface and a light source at infinity with angle 𝜃 .

𝜃 𝜃

−𝜃 𝜋

2 𝜋

2

Assume a box BRDF, flat surface and a light source at infinity with angle 𝜃 .

𝜃 * = 𝜃

−𝜃 𝜋

2

−𝑎 𝑎 𝜋

2 x (space) x (space)

Reflection - BRDF

Spatial domain:

The BRDF action on a light field is a convolution with the BRDF function 𝑙

𝑅′

= 𝑙

𝑅

∗ 𝜌(𝜃)

Frequency domain:

Convolution is changed into pointwise multiplication 𝑙

𝑅′

Ω 𝑥

, Ω 𝜃

= 𝑙

𝑅

Ω 𝑥

, Ω 𝜃

𝐹(𝜌(Ω 𝜃

))

BRDF example

Type equation here.

Incoming

Light field

⊗ Light field

After BRDF change

× -

Download