Szeliski Being There

Being There:Capturing and Experiencing aSense of Place: 

Being There: Capturing and Experiencing a Sense of Place Richard Szeliski Microsoft Research Symposium on Computational Photography and Video

Early art: events: 

Early art: events Lascaux

Early art: events: 

Early art: events

Early art: events: 

Early art: events Bayeux Tapestries

Renaissance art: events: 

Renaissance art: events Da Vinci

Baroque art: portraits: 

Baroque art: portraits Rembrandt

Baroque art: landscapes: 

Baroque art: landscapes Rembrandt

Classical art: landscapes: 

Classical art: landscapes Corot

Impressionism: landscapes: 

Impressionism: landscapes Monet

Modern art: multi-perspective: 

Modern art: multi-perspective Picasso

Impressionism: landscapes: 

Impressionism: landscapes Monet

Photography: events: 

Photography: events Henri Cartier-Bresson

Photography: landscapes: 

Photography: landscapes Ansel Adams

Photography: landscapes: 

Photography: landscapes Galen Rowell

Photography: landscapes: 

Photography: landscapes Art Wolfe

Interactive Visual Media: 

Interactive Visual Media Branching storytelling Games: problem solving, skill acquisition sense of place, navigation

Real-world “immersive” media: 

Real-world “immersive” media Recapturing the sense of place QTVR: the first real breakthrough in modeling from reality? Fairly universal for high-end home and car sales, virtual tourism Limitations: mostly static, discrete jumps

Increasing realism: 

Increasing realism Better: field of view and resolution Demo: Space Needle dynamic range Additional cues/modalities: motion / movement Demo: Deception Falls sound

VR Tools: 

VR Tools

Fully automated stitch discovery: 

Fully automated stitch discovery Demo

Image stitching — open issues: 

Image stitching — open issues Fully automated assembly (no ordering) Full 2D stitching (multiple rows) Double image fix-up (de-ghosting) Merging different exposures Automated grouping/clustering (detection)

HDR: merging exposures: 

HDR: merging exposures Inputs Tonemapped output: no motion compensation or consistency check

HDR: merging exposures: 

Inputs Tonemapped output: global+local motion compensation HDR: merging exposures

Video Textures: 

Video Textures How can we turn a short video clip into continuous video? Enhance sense of liveness Use in games and presentations

Video Textures: 

Video Textures Find cyclic structure in the video Play frames with random shuffle Optional region-based analysis Smooth over discontinuities (morph)

Animating Stills: 

Animating Stills What if we only have a single beautiful photo (or painting)? Can we add so liveness to a photo/still?

Animating Stills: 

Animating Stills

Increasing “immersion”: 

Increasing “immersion” Add continuous movement

DemoBellevue Botanical Garden: 

Demo Bellevue Botanical Garden

DemoHDR Home Walkthrough: 

Demo HDR Home Walkthrough

Video-based walkthroughs: 

Video-based walkthroughs Add continuous movement How did we do this? Is this 3D? 2D? Graphics? Video? Dimensionality and storage?

Limits of Video-Based Walkthroughs: 

Limits of Video-Based Walkthroughs Video-based walkthroughs are “rail experiences” of a space-time slice How do we break the bounds? Can’t move from side to side If there is motion in the scene, don’t see it Capturing all points of view everywhere in space at all times is impractical

A Practical 3D Video Camera: 

A Practical 3D Video Camera Working volume? Walls of a room: Virtualized Reality 2D “window”: LightField Array 1D “rail”: Virtual Viewpoint Video

“3D” video: 

“3D” video

Virtual Viewpoint Video: 

Virtual Viewpoint Video Capture multiple synchronized video streams

Key to view interpolation: Geometry: 

Key to view interpolation: Geometry Stereo Camera 1 Camera 2 Image 1 Image 2 Virtual Camera

Don’t match pixels – match segments: 

Don’t match pixels – match segments Segments contain more information, so they’re easier to match.

Iteratively update depths: 

Iteratively update depths

Depth through time:: 

Depth through time:

Matting: 

Matting Some pixels get influence for multiple surfaces. Background Surface Foreground Surface Image Camera Close up of real image: Multiple colors and depths at boundary pixels…

Find matting information:: 

Find matting information: Find boundary strips using depth. 2. Within boundary strips compute the colors and depths of the foreground and background object. Background Foreground Strip Width

Why matting is important: 

Why matting is important Matting No Matting

Representation: 

Representation Color Depth MAIN LAYER Color Depth Matting information BOUNDARY LAYER

Rendering pipeline (GPU): 

Rendering pipeline (GPU) Render Main Layer Render Boundary Layer Composite Render Main Layer Render Boundary Layer

Real-time viewpoint control: 

Real-time viewpoint control “Matrix”-like effects – on demand! (With much, much less hardware.)

Wrapping up…: 

Wrapping up…

Light Field Rendering: 

Light Field Rendering Sample and synthesize (capture and render) from a 4D ray space [Lightfield, Levoy & Hanrahan; Lumigraph, Gortler et al., SIGGRAPH’96]

Slow Glass: 

Slow Glass What if we could capture all the photons and play them all back at a later time? Light of Other Days Bob Shaw, © 1969 http://www.scifi.com/scifiction/classics/classics_archive/shaw/shaw1.htm

Slow Glass: 

Slow Glass One could stand the glass beside, say, a woodland lake until the scene emerged, perhaps a year later. If the glass was then removed and installed in a dismal city flat, the flat would—for that year—appear to overlook the woodland lake. During the year it wouldn't be merely a very realistic but still picture—the water would ripple in sunlight, silent animals would come to drink, birds would cross the sky, night would follow day, season would follow season. Apart from its stupendous novelty value, the commercial success of slow glass was founded on the fact that having a scenedow was the exact emotional equivalent of owning land. The meanest cave dweller could look out on misty parks—and who was to say they weren't his? A man who really owns tailored gardens and estates doesn't spend his time proving his ownership by crawling on his ground, feeling, smelling, tasting it. All he receives from the land are light patterns…

Being There: 

Being There Artists (and all of us) have always wanted to capture a sense of being there. Computational photography and video bring us a lot closer: realism (field of view, resolution, contrast) movement and sound immersion and exploration What does the future hold?