Introduction Thomas Burnett CTO, founder, and primary investigator - - PowerPoint PPT Presentation

GTC 2018

FoVI3D

Extreme Multi-view Rendering for Light-field Displays

Introduction

2

Thomas Burnett

CTO, founder, and primary investigator for FoVI3D. ~15 years experience developing rendering solutions and architectures for static and dynamic light-field display systems. tburnett@fovi3D.com

Strong technical team with deep experience in optical, mechanical, electrical and software engineering.

Agenda

3

• Definition and Significance of

Light-field Displays

• Extreme Multi-view Rendering

for Light-Field Displays

4

Definition and Significance of Light-field Displays

Light-field Displays in Popular Media

5

Light-field Display Definition

6

• Reproduces a 3D aerial image visible to the unaided eye without glasses or

head tracking

• Binocular disparity, occlusion, specular highlights, and gradient shading and
• ther expected depth cues must be correct from the viewer’s perspective as in

the natural real-world light-field

Significance of Light-field Displays

7

• Human binocular vision and acuity, and the

accompanying 3D retinal processing of the human eye and brain, are specifically designed to promote situational awareness and understanding in the natural 3D world.

• The ability to resolve depth within a scene,

whether natural or artificial, improves our spatial understanding of the environment and as a result reduces the cognitive load accompanying the analysis and collaboration on complex tasks.

Natural vs Synthetic Light-fields

8

In a natural light-field we observe light reflected

• ff objects, perceiving color and depth cues. An

eye focuses a 3D scene on the retina as a 2D image, our brain reconstructs the 3D world. In a synthetic light-field, we observe light projected from a surface that generates the same color and depth cues, thus we see the same object.

9

Extreme Multi-view Rendering for Light-Field Displays

What is a Light-field

10

• A light-field is a set of rays that pass through every point in space and is typically defined for computer vision by a

plenoptic function: L = P Θ, ϕ, λ, Vx, Vy, Vz

• Can be presented as a 2D raster image (Radiance Image) where each pixel represents the color, position, and

direction of a ray within the light-field

• A light-field display is essentially a plenoptic projector; the inverse of a plenoptic camera

A hogel (holographic element) is a radiance image for a single micro-lens. Light-field displays require many hogels and thus extremely large radiance images.

Light-field Display Resolution

11

Spatial Resolution: Determined by size and pitch of micro lenses. Higher density, smaller micro-lens gives higher spatial resolution. Directional Resolution (Dr): Determined by the dimensions of the hogel. Assuming a square hogel, Dr2 is the number of rays per micro-lens. More is better. Field of View (FoV) The angular projection of a micro- lens. Angular Pitch (Ap): FoV/Dr. More than one ray per degree is desirable.

FoV

Dr

Magnitude of the Light-field Radiance Image

Size of Radiance Image - Examples

• 1.0mm hogels, Dr = 256, ~300mm (1ft) usable depth
• 0.7mm hogels, Dr = 512, ~600mm (2ft) usable depth
• 0.5mm hogels, Dr = 128, ~150mm (6”-8”) usable depth – real time 30 fps rendering

1,429 hogels x 1,071 hogels x 256 rays x 256 rays x 3 RGB bytes per pixel = ~150 Gigabytes per frame 1,429 hogels x 1,071 hogels x 512 rays x 512 rays x 3 RGB bytes per pixel = ~1.2 Terabytes per frame 2,000 hogels x 1,500 hogels x 128 rays x 128 rays x 3 RGB bytes per pixel = ~150 Gigabytes per frame x 30 fps = ~4.4 Terabytes per second

1m x 0.75m (90° FoV) Static Light-field TopoMaps for US Military 1m x 0.75m (90° FoV) Dynamic, Real Time, Light-field Display Conclusion: DO NOT MOVE PIXELS, don’t transport them, don’t store them. Render in hardware at the display instead.

Light-field Radiance Image

13

3D Model with Hogel Image Plane Radiance Image

502 Hogels – 402 Pixels (”Rays”) per Hogel

Resultant 3D Aerial LfD Projection

Cluster Based Radiance Image Rendering

14

Host Application Scene

Light-field Display interactivity and update rate are proportional to the complexity of the scene/model, the power/configuration of the rendering cluster and the size of the light-field display radiance image.

The OpenGL Shim

15

GL Application GL Shim Opengl32.dll OpenGL Driver Graphics Card GPU Dynamic Display Driver

DVI Ethernet

• 1. An OpenGL Shim is an ‘interceptor’

library design to mimic and replace/forward graphics commands.

• 2. In essence, the OpenGL shim

highjacks the command stream unbeknownst to the host application.

• 3. Often accompanied by a display

specific companion library.

Zebra Imaging ZScape Motion Display

Light-field Display Developed by Zebra Imaging under the DARPA UPSD Program

https://youtu.be/blb0TUBoZwA https://youtu.be/b_CKQN1t-e8

OpenGL and GPUs for Light-field Rendering

16

Many developers of novel display architectures have (or had) developed fix function OpenGL shims to intercept draw commands and forward them to a display specific renderer.

• However, there is no agreement on which OpenGL version and/or functions to shim.
• There are legacy fixed function APIs which are suboptimal for specifying geometry that programmers often default to because of their ease of

use and readability. These immediate mode functions perform poorly in the context of multi-view rendering. Their counterpart functions, the retained mode APIs, are far better suited for multi-view rendering; however, the retained model APIs have evolved significantly over the course of the OpenGL specification and not every application is implemented with the best constructs.

• Many of the OpenGL functions and functionality are not applicable to distributed or multi-view rendering. Point and line rendering is

computed in screen space and is not 3D.

• Shaders make shimming with confidence nearly impossible or is so highly restricted to the point where shaders may just emulate a fixed

function pipeline with little variation.

• There usually exists additional display specific extensions or APIs which then reduce the portability even more.

OpenGL has a single active viewport and a single active view matrix

• Fixed function OpenGL defines a single GL_MODELVIEW matrix which is the pre-multiplied model and view matrices. This implies that for any

display that requires multi-view rendering, the view component needs to be removed from the GL_MODELVIEW matrix or that the view matrix is specified in some other manner.

• Multi-viewpoint rendering becomes a responsibility of the host application which must cache the render commands and then re-render the list

for each viewpoint.

• Culling operations may have to be disabled since the host application may not understand the viewpoints or projections necessary for the

display technology.

• Meanwhile, the host application has stalled until all the views are rendered because the render state can not change until all the views are

rendered. The application developer has to know a lot about the target display.

• Number of views
• View specific projections and transforms
• Distortion shaders

Bounding Volumes Example

17

Double Frustum Rendering

18

• Advantage of rendering a hogel natively
• Does require two passes of the geometry to

create the ‘double frustum’ and there can be a mathematical singularity at the center of the bowtie

• Thousand of render passes per display update
• Very small viewport
• At least within the fixed-function OpenGL render

pipeline, camera matrix cannot have a non-zero near plane

• Benefits from frustum culling is tightly bound to

a hogel’s projection frustum

• Benefits from small batch size

Halle, Michael W. “Fast Computer Graphics Rendering for Full Parallax Spatial Displays.” International Society for Optics and Photonics.

Double Frustum Near Clip Planes

19

Oblique Slice and Dice

20

• Typically makes better use of the GPU in

that the algorithm generally renders to larger framebuffers within GPU memory

• May require fewer passes of the

geometry than the double frustum algorithm, but still hundreds to thousand of render passes

• Rendered pixels are not in a spatial form

that can be projected through a micro- lens array

• Oblique frustum is highly likely to

intersect changed geometry whether or not an individual hogel frustum actually intersects that geometry

• Bounding volumes have less impact on

the oblique rendering rates

• Batch size as lesser effect
• May require sampling if the hogel

arrangement is not rectilinear grid

• Inhibits on-the-fly transform

Current Graphics Rendering: The Tightly Bound Display

21

Server

3D Data Service

Client System

Data Video

Display 3D App

Render Engine GPU

Draw Commands

Today’s Tightly Bound Display Tomorrow’s Heterogeneous Display Ecosystem Server

3D Data Service

Client System

Data ObjGL

Display System 3D App

Scene Geometry

MvPU MvPU MvPU

…

App is agnostic to display architecture

Tomorrow’s Display Environment is Heterogenous

22

HDE: Sport Visualization

23

Flat Panel: PoV 0 Flat Panel PoV 4 Light-field Display: Bird’s Eye View 0 Flat Panel: PoV Stat 0 Beer HMD PoV 2

Light-field Display Architecture

24

Light-field Display Developer Kit

25

Lab Prototype Display With Bring-up Lenses LFD Developer Kit with Final Lens Solution First Quarter 2018 LfD DK2 https://youtu.be/oRjKKJM8IQc