1 Linear View Synthesis Using a Dimensionality Gap Light Field Prior Anat Levin and Fredo Durand Weizmann Institute of Science & MIT CSAIL Light fields Light field: the set of rays emitted from a scene in all possible directions 2 2 Light fields Novel view rendering (Animation by Marc Levoy) 3 3 Light fields Novel view rendering (Animation by Marc Levoy) 4 4 Light fields Novel view rendering (Animation by Marc Levoy) Synthetic refocusing 5 5 4D light field v The set of light rays hitting the camera aperture plane is 4D: • Ray hitting point- 2D • Ray orientation- 2D u (In general: a 7D plenoptic space, including time and wavelength dimensions) 6 6 Light field acquisition schemes and priors 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 7 7 Light field acquisition schemes and priors • 4D: The light field is smooth, but involves 4 degrees of freedom -Capture: 4D data (e.g. camera array) -Inference: linear 8 8 Light field acquisition schemes and priors • 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 9 9 In this talk: 3D light field prior 10 10 • 4D: -Capture: 4D data (e.g. camera array -Inference: linear • 2D: -Capture: 2D data (e.g. stereo camera) y -Inference: non linear depth estimation v u x • 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) -Inference: linear Outline • Linear view synthesis from a focal stack sequence • The 3D light field prior • Frequency derivation of synthesis algorithm • Other applications of the 3D prior 11 11 Linear view synthesis with 3D prior 12 12 Input: Focal stack (3D data) Output: Novel viewpoints (4D data) 1D set of 2D images focused at different depth 2D Images x 2D set of novel viewpoints Linear image processing 13 13 Linear view synthesis algorithm No depth estimation! Shift focal stack images by disparity of desired view 1 Average shifted images 2 Depth invariant deconvolution 3 14 14 Shift invariant convolution~ focus sweep camera Average shifted images Depth invariant blur kernel 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 Ideal pinhole image Linear view synthesis results Video animation here 15 15 Disclaimers • 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 16 16 Outline • Linear view synthesis from a focal stack sequence • The 3D light field prior • Frequency derivation of synthesis algorithm • Other applications of the 3D prior 17 17 4D light field v y v 18 y x u x • The set of light rays hitting the lens is 4D u (x,y,u,v) 4D light field v y v 19 y u x x (?,?,u0,0) • The set of light rays hitting the lens is 4D (x,y,u,v) u 4D light field 20 v y v y x u x u (?,?,0,v0) • The set of light rays hitting the lens is 4D (x,y,u,v) 4D light field v y v 21 y x u x • The set of light rays hitting the lens is 4D u (x,y,u,v) 4D light field spectrum 22 y y v u v x x 4D Fourier Transform • The set of light rays hitting the lens is 4D • Study the 4D Fourier domain u (x,y,u,v) L( x, y, u,v) 4D light field spectrum 23 y y v v u v u x x 4D Fourier Transform L( x0,0,?,?) • The set of light rays hitting the lens is 4D • Study the 4D Fourier domain u (x,y,u,v) L( x, y, u,v) 4D light field spectrum 24 y y v u x 4D Fourier Transform Frequency content only along 1D segments v u x 4D light field spectrum Scene 4D Light field spectrum Energy portion away from focal segments The slicing theorem 26 y y v u v u x x 4D Fourier Transform 2D focused images at varying depths 2D Fourier Transform The dimensionality gap 27 y y v u near x far 4D Fourier Transform Light field spectrum: 4D Image spectrum: 2D 3D Depth: 1D → Dimensionality gap (Ng 05, Levin et al. 09) Only the 3D manifold corresponding to physical focusing distance is useful vv uu x 3D Gaussian light field prior 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 y 28 v u x Outline • Linear view synthesis from a focal stack sequence • The 3D light field prior • Frequency derivation of synthesis algorithm • Other applications of the 3D prior 29 View synthesis in the frequency domain Average focal stack spectra Spectra of Sample correct depth density 1 30 4D spectrum of y constant depth v scene u x Deconvolution (frequency domain) Spectra of focal stack images Outline • Linear view synthesis from a focal stack sequence • The 3D light field prior • Frequency derivation of synthesis algorithm • Other applications of the 3D prior 31 Prior to infer light field from partial samples In many other light field acquisition schemes we capture only a partial information on the light fieldlimited 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. 32 Improved viewpoints sample 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 Circle: Sampling pattern with sampling pattern improved reconstruction using 3D prior 33 Superesolution of plenoptic camera measurements34 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 measurements35 Bicubic interpolation Our result: applies for all depths simultaneously, no depth estimation Lumsdaine and Georgiev: applies for a single known depth Summary • Light field acquisition and synthesis strongly depends on light field prior Existing priors: 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 • Linear view synthesis from the focal stack • Other applications of 3D prior: - viewpoints sample pattern - depth invariant superesolution of plenoptic camera data 36