Linear View Synthesis Using a Dimensionality Gap Light Field Prior

Slides:



Advertisements
Similar presentations
Bayesian Belief Propagation
Advertisements

A Robust Super Resolution Method for Images of 3D Scenes Pablo L. Sala Department of Computer Science University of Toronto.
Fourier Slice Photography
Light Fields PROPERTIES AND APPLICATIONS. Outline  What are light fields  Acquisition of light fields  from a 3D scene  from a real world scene 
Light Field Rendering Shijin Kong Lijie Heng.
Light Field Stitching with a Plenoptic Camera Zhou Xue LCAV - École Polytechnique Fédérale de Lausanne Dec

BMME 560 & BME 590I Medical Imaging: X-ray, CT, and Nuclear Methods Tomography Part 3.
Light Mixture Estimation for Spatially Varying White Balance
Digital Image Processing Chapter 4: Image Enhancement in the Frequency Domain.
lecture 5, Sampling and the Nyquist Condition Sampling Outline  FT of comb function  Sampling Nyquist Condition  sinc interpolation Truncation.
Midterm Review. World is practically continuous in time space color brightness dynamic range brightness.
Advanced Computer Graphics (Fall 2010) CS 283, Lecture 17: Frequency Analysis and Signal Processing for Rendering Ravi Ramamoorthi
CSCE641: Computer Graphics Image Formation Jinxiang Chai.
CSCE 641 Computer Graphics: Image-based Rendering (cont.) Jinxiang Chai.
Comp 665 Convolution. Questions? Ask here first. Likely someone else has the same question.
Computational Photography Light Field Rendering Jinxiang Chai.
Frequency Analysis and Sheared Reconstruction for Rendering Motion Blur Kevin Egan Yu-Ting Tseng Nicolas Holzschuch Frédo Durand Ravi Ramamoorthi Columbia.
Rendering with Concentric Mosaics Heung-Yeung Shum Li-Wei he Microsoft Research.
CSCE 641: Computer Graphics Image-based Rendering Jinxiang Chai.
Understanding and evaluating blind deconvolution algorithms
The Story So Far The algorithms presented so far exploit: –Sparse sets of images (some data may not be available) –User help with correspondences (time.
lecture 2, linear imaging systems Linear Imaging Systems Example: The Pinhole camera Outline  General goals, definitions  Linear Imaging Systems.
Basic Principles of Imaging and Photometry Lecture #2 Thanks to Shree Nayar, Ravi Ramamoorthi, Pat Hanrahan.
Multi-Aperture Photography Paul Green – MIT CSAIL Wenyang Sun – MERL Wojciech Matusik – MERL Frédo Durand – MIT CSAIL.
Light Field. Modeling a desktop Image Based Rendering  Fast Realistic Rendering without 3D models.
Advanced Computer Graphics CSE 190 [Spring 2015], Lecture 3 Ravi Ramamoorthi
1 Fabricating BRDFs at High Spatial Resolution Using Wave Optics Anat Levin, Daniel Glasner, Ying Xiong, Fredo Durand, Bill Freeman, Wojciech Matusik,
Echivalarea sistemelor analogice cu sisteme digitale Prof.dr.ing. Ioan NAFORNITA.
Antialiasing CptS 548 Advanced Computer Graphics John C. Hart.
1 Patch Complexity, Finite Pixel Correlations and Optimal Denoising Anat Levin, Boaz Nadler, Fredo Durand and Bill Freeman Weizmann Institute, MIT CSAIL.
Image Formation. Input - Digital Images Intensity Images – encoding of light intensity Range Images – encoding of shape and distance They are both a 2-D.
Introduction to Computational Photography. Computational Photography Digital Camera What is Computational Photography? Second breakthrough by IT First.
01/28/05© 2005 University of Wisconsin Last Time Improving Monte Carlo Efficiency.
Automatic Registration of Color Images to 3D Geometry Computer Graphics International 2009 Yunzhen Li and Kok-Lim Low School of Computing National University.
Dynamically Reparameterized Light Fields Aaron Isaksen, Leonard McMillan (MIT), Steven Gortler (Harvard) Siggraph 2000 Presented by Orion Sky Lawlor cs497yzy.
Computational photography CS4670: Computer Vision Noah Snavely.
Image-based rendering Michael F. Cohen Microsoft Research.
Correspondence-Free Determination of the Affine Fundamental Matrix (Tue) Young Ki Baik, Computer Vision Lab.
1 Finding depth. 2 Overview Depth from stereo Depth from structured light Depth from focus / defocus Laser rangefinders.
Image Processing Edge detection Filtering: Noise suppresion.
03/24/03© 2003 University of Wisconsin Last Time Image Based Rendering from Sparse Data.
Fourier Analysis of Stochastic Sampling For Assessing Bias and Variance in Integration Kartic Subr, Jan Kautz University College London.
1 Plenoptic Imaging Chong Chen Dan Schonfeld Department of Electrical and Computer Engineering University of Illinois at Chicago May
Fourier Depth of Field Cyril Soler, Kartic Subr, Frédo Durand, Nicolas Holzschuch, François Sillion INRIA, UC Irvine, MIT CSAIL.
CSL 859: Advanced Computer Graphics Dept of Computer Sc. & Engg. IIT Delhi.
Panorama artifacts online –send your votes to Li Announcements.
112/5/ :54 Graphics II Image Based Rendering Session 11.
FREE-VIEW WATERMARKING FOR FREE VIEW TELEVISION Alper Koz, Cevahir Çığla and A.Aydın Alatan.
 Marc Levoy Using Plane + Parallax to Calibrate Dense Camera Arrays Vaibhav Vaish, Bennett Wilburn, Neel Joshi, Marc Levoy Computer Science Department.
Extracting Depth and Matte using a Color-Filtered Aperture Yosuke Bando TOSHIBA + The University of Tokyo Bing-Yu Chen National Taiwan University Tomoyuki.
Lecture#4 Image reconstruction
CSCE 641 Computer Graphics: Image-based Rendering (cont.) Jinxiang Chai.
1Ellen L. Walker 3D Vision Why? The world is 3D Not all useful information is readily available in 2D Why so hard? “Inverse problem”: one image = many.
Image-Based Rendering Geometry and light interaction may be difficult and expensive to model –Think of how hard radiosity is –Imagine the complexity of.
Announcements Final is Thursday, March 18, 10:30-12:20 –MGH 287 Sample final out today.
Today Defocus Deconvolution / inverse filters. Defocus.
RECONSTRUCTION OF MULTI- SPECTRAL IMAGES USING MAP Gaurav.
Presented by 翁丞世  View Interpolation  Layered Depth Images  Light Fields and Lumigraphs  Environment Mattes  Video-Based.
Multi-Aperture Photography
Extended Depth of Field For Long Distance Biometrics
- photometric aspects of image formation gray level images
Echivalarea sistemelor analogice cu sisteme digitale
FFT-based filtering and the
Motion and Optical Flow
Deconvolution , , Computational Photography
Computer Vision Lecture 4: Color
Computational Plenoptic Imaging
Filtering Things to take away from this lecture An image as a function
Filtering An image as a function Digital vs. continuous images
Presentation transcript:

Linear View Synthesis Using a Dimensionality Gap Light Field Prior Anat Levin and Fredo Durand Weizmann Institute of Science & MIT CSAIL 1

Light fields 2 Light field: the set of rays emitted from a scene in all possible directions

(Animation by Marc Levoy) Light fields 3 Novel view rendering (Animation by Marc Levoy)

(Animation by Marc Levoy) Light fields 4 Novel view rendering (Animation by Marc Levoy)

(Animation by Marc Levoy) Light fields 5 Novel view rendering Synthetic refocusing (Animation by Marc Levoy)

4D light field 6 u v The set of light rays hitting the camera aperture plane is 4D: Ray hitting point- 2D Ray orientation- 2D (In general: a 7D plenoptic space, including time and wavelength dimensions)

Light field acquisition schemes and priors 7 Very different approaches to light field acquisition and manipulations exist in the literature. The inherent difference between them is a different prior model on the light field space

Light field acquisition schemes and priors 8 4D: The light field is smooth, but involves 4 degrees of freedom -Capture: 4D data (e.g. camera array) -Inference: linear .

Light field acquisition schemes and priors 9 4D: -Capture: 4D data (e.g. camera array) -Inference: linear 2D: For Lambertian scenes all rays emerging from one point have same color. If depth is known, only 2 degrees of freedom -Capture: 2D data (e.g. stereo camera) -Inference: non linear depth estimation

In this talk: 3D light field prior 10 4D: -Capture: 4D data (e.g. camera array -Inference: linear 2D: -Capture: 2D data (e.g. stereo camera) -Inference: non linear depth estimation 3D: Depth is a 1D variable, hence the union of images at any depth covers no more than a 3D subset. Show that in the frequency domain there is only a 3D manifold of non zero entries. -Capture: 3D data (e.g. focal stack)

Outline Linear view synthesis from a focal stack sequence 11 Linear view synthesis from a focal stack sequence The 3D light field prior Frequency derivation of synthesis algorithm Other applications of the 3D prior

Linear view synthesis with 3D prior 12 Input: Focal stack (3D data) 1D set of 2D images focused at different depth Output: Novel viewpoints (4D data) 2D Images x 2D set of novel viewpoints Linear image processing

Linear view synthesis algorithm 13 No depth estimation! Shift focal stack images by disparity of desired view Average shifted images Depth invariant deconvolution 1 2 3

Shift invariant convolution~ focus sweep camera 14 Average shifted images Depth invariant blur kernel Ideal pinhole image Inspiration: The focus sweep camera Hausler 72, Nagahara et al. 08 Captures a single image, average over all focus depths during exposure, provides EDOF image from a single view

Linear view synthesis results 15 Video animation here

Disclaimers Novel viewpoints limited to the aperture area 16 Novel viewpoints limited to the aperture area Convolution model breaks at occlusion boundaries Assume scene is Lambertian- in practice holds within the narrow range of angles of the aperture

Outline Linear view synthesis from a focal stack sequence 17 Linear view synthesis from a focal stack sequence The 3D light field prior Frequency derivation of synthesis algorithm Other applications of the 3D prior

4D light field v y v x u u (x,y,u,v) 18 u v u v y x y x (x,y,u,v) The set of light rays hitting the lens is 4D

4D light field v y v x u u (?,?,u0,0) (x,y,u,v) 19 u v u v y x y x (?,?,u0,0) (x,y,u,v) The set of light rays hitting the lens is 4D

4D light field v y v x u u (?,?,0,v0) (x,y,u,v) 20 u v u v y x y x (?,?,0,v0) (x,y,u,v) The set of light rays hitting the lens is 4D

4D light field v y v x u u (x,y,u,v) 21 u v u v y x y x (x,y,u,v) The set of light rays hitting the lens is 4D

4D light field spectrum y v x u (x,y,u,v) L( , , , ) 22 v y x u 4D Fourier Transform (x,y,u,v) The set of light rays hitting the lens is 4D Study the 4D Fourier domain L( , , , )

4D light field spectrum y v x u L( ,0,?,?) (x,y,u,v) L( , , , ) 23 v y x u 4D Fourier Transform L( ,0,?,?) (x,y,u,v) The set of light rays hitting the lens is 4D Study the 4D Fourier domain L( , , , )

4D light field spectrum y v x u 24 v y x u 4D Fourier Transform Frequency content only along 1D segments

Energy portion away from focal segments 4D light field spectrum Scene 4D Light field spectrum Energy portion away from focal segments

The slicing theorem y v x u 4D Fourier Transform 26 v y x u 4D Fourier Transform 2D focused images at varying depths 2D Fourier Transform

The dimensionality gap 27 v y x u far 4D Fourier Transform depth color coding near Light field spectrum: 4D Image spectrum: 2D Depth: 1D → Dimensionality gap (Ng 05, Levin et al. 09) Only the 3D manifold corresponding to physical focusing distance is useful 3D

3D Gaussian light field prior 28 Gaussian prior: assigns non zero variance only to 3D set of entries on the focal segments Gaussian=> inference simple and linear Focal stack directly samples the manifold with non zero variance

Outline Linear view synthesis from a focal stack sequence 29 Linear view synthesis from a focal stack sequence The 3D light field prior Frequency derivation of synthesis algorithm Other applications of the 3D prior

View synthesis in the frequency domain 30 4D spectrum of constant depth scene Average focal stack spectra Spectra of correct depth Sample density Spectra of focal stack images Deconvolution (frequency domain)

Outline Linear view synthesis from a focal stack sequence 31 Linear view synthesis from a focal stack sequence The 3D light field prior Frequency derivation of synthesis algorithm Other applications of the 3D prior

Prior to infer light field from partial samples 32 In many other light field acquisition schemes we capture only a partial information on the light field- limited resolution, aliasing and each. However, we capture linear measurements On the other hand, we have a Gaussian prior, and we know the light field actually occupies only a low dimensional manifold of the 4D space. Use the prior to “invert the rank deficient projection” and interpolate the measurements to get a light field with higher resolution, less aliasing.

Improved viewpoints sample 33 4D Light field acquisition systems sample a 2D set of view points Can we do with sparser sample and 3D Gaussian prior for interpolation? How many samples needed? What is the right spacing? Shall we distribute samples on a grid? Better arrangement? Grid: Standard sampling pattern Circle: Sampling pattern with improved reconstruction using 3D prior

Superesolution of plenoptic camera measurements 34 Plenoptic camera measurements are aliased Replicas off the focal segments are high frequencies which we can re-bin and restore high frequency information

Superesolution of plenoptic camera measurements 35 Bicubic interpolation Lumsdaine and Georgiev: applies for a single known depth Our result: applies for all depths simultaneously, no depth estimation

Summary 36 Light field acquisition and synthesis strongly depends on light field prior Existing priors: Linear view synthesis from the focal stack Other applications of 3D prior: - viewpoints sample pattern - depth invariant superesolution of plenoptic camera data 4D prior: capture- 4D data (e.g. camera array), inference- linear 2D prior: capture- 2D data (e.g. stereo), inference- non linear Our new prior: 3D prior: capture- 3D data (e.g. focal stuck), inference linear