Emerging Convergence of Video, Vision & Graphics
Harpreet S. Sawhney
Rakesh Kumar

ACKNOWLEDGEMENTS
Collaborative Work with:
Hai Tao
Yanlin Guo
Steve Hsu
Supun Samarasekera
Keith Hanna
Aydin Arpa
Rick Wildes

TECHNICAL SUCCESS OF CONVERGENCE
TECHNOLOGIES
PC based near real-time mosaicing
Image based modeling for Entertainment
Automated Video Enhancement: VHS-to-DVD
Real-time Video Insertion
Iris recognition, active vision

Immersive and Interactive Telepresence
Modes of Operation
Observation Mode
User observes a remote site from any perspective.
User “walks” through site to view activities of interest
“up close”.
Example: security, facility guards, sports & entertainment
Conversation Mode
Users talk and observe one another as if in the same room.
Users walk around yet maintain eye contact.
Example: immersive teleconferencing
Interaction Mode
Remote users share a common work space.
Users observe each other’s hands as they manipulate shared objects, such as war room wall displays.
Example: mission planning, remote surgery

Quality of Service for Tele-presence
Critical Issues
• High quality for immersive experience
– Artifact free recovery of 3D shape from video streams
– Efficient 3D video representation and compression
– High quality rendering of new views using 3D shape and video streams
– Bandwidth available in the Next Generation Internet
• Low latency for interactive applications
– Real time 3D geometry recovery at the content server end
– Real time new view rendering at the browser client end
– Adaptive Stream management to handle user requests and network loads
– Error resilience and concealment to fill in missing packets

Convergence Technologies
… for immersive & interactive visual applications ...
• Vision algorithms: High-quality 3D shape recovery and dynamic scene analysis
• ASICs, high performance hardware: Real-time video processing
• Compact, low-cost cameras: CMOS cameras
• Low latency and high quality compression: Error resilience
• Real time view synthesis : Standard platforms, e.g. PCs
• Immersive Displays

Vision algorithm performance over time
High Quality 3d shape extraction
2000
Video registration to 3D site models
1998
Coarse 3D Depth Recovery
1995
2D Video Insertion
1993
2D Stabilization
1990
Immersive
Telepresence
Georegistration visual databases
Face Finding for
Iris Recognition
Real-time insertion in
Live TV
Mosaicing for entertainment & surveillance
Time

VFE-100
1992
VFE-200
1997
ACADIA ASIC
2000
• Sarnoff ACADIA ASIC performance
• 100 MHz system clock, processes 100 million pixels/sec in each processing element
• 10 billion operations / sec total IC performance
• 800 MB/sec SDRAM interface using 64-bit bus
• Enables building smart 3D cameras for immersive applications.

Application Performance
• Parametric Motion : Stabilization & Mosaicing
– 720x240 fields @ 60 Hz OR 720x480 frames @ 30 Hz
• Pyramid based Fusion : Dynamic Range, Focus
Enhancement
– 720x240 fields @ 60 Hz OR 720x480 frames @ 30 Hz
• Stereo Depth Extraction
– 720x240 field 32 disparity levels in 4 ms (250 Hz)
– 720x240 field 60 disparity levels in 10 ms (100 Hz)
– 60 disparities on 1k x 1k images at 55 ms (18 Hz)

… Required algorithm components for tele-presence are emerging ...
MPEG4, Progressive Encoding
1999
E-vue
Low Latency MPEG2 multiplexing service
1998-1999
Just Noticeable Difference (JND):
MPEG2 Encoding and Quality
Measurement
1997-1998
Tektronix
VideoPhone: H.263
1997-1998
LG Electronics
MPEG2: Encoding and Transmission
1993- 1996
DIREC-TV & HDTV
Pyramid & Wavelet based Encoding
1988-1993
Still Image Compression
Time
ICTV

A FRAMEWORK FOR VIDEO PROCESSING
ALIGN
2D & 3D MODELS OF MOTION & STRUCTURE
MODEL-BASED IMAGE SEQUENCE ALIGNMENT
TEST
WARP/RENDER WITH 2D/3D MODELS
TEST ALIGNMENT QUALITY
SYNTHESIZE
CREATE OUTPUT REPRESENTATIONS

Highlights of Sarnoff’s Video Analysis Technologies
… framework applied to a create immersive representations ...
2D Immersive
& Layered Representations
Spherical Mosaics
Dynamic & Synopsis Mosaics
Core Vision Algorithms for (Real-time)
Motion & 3D Video Analysis
Model-centric
Video Visualization
Dynamic model & video visualization
Geo-registration with reference image database
Stereo & Video Sequence
Enhancement
Hi-Q IBR based mixed resolution synthesis
Video Quality Enhancement for efficient compression
Multi-camera Immersive
Dynamic Rendering
Hi-Q Depth extraction
Image-based rendering with dynamic depth

TOPOLOGY INFERENCE & LOCAL-TO-GLOBAL ALIGNMENT
SPHERICAL MOSAICS
[Sawhney,Hsu,Kumar ECCV98, Szeliski,Shum SIGGRAPH98]
Sarnoff Library Video
Captures almost the complete sphere with 380 frames

SPHERICAL TOPOLOGY EVOLUTION

SPHERICAL MOSAIC
Sarnoff Library

ACTIVE FOCUS OF ATTENTION
WFOV/NFOV CONTROL

Original Video
DYNAMIC MOSAICS
Video Stream with deleted moving object
Dynamic Mosaic Video

SYNOPISIS MOSAICS

ALIGNMENT & SYNTHESIS FOR HI-RES STEREO SYNTHESIS
A HIGH END APPLICATION OF IBMR
[Sawhney,Guo,Hanna,Kumar,Zhou,Adkins SIGGRAPH2001]
Low-Res Left
Original High-Res Right
Synthesized High-Res Left

Left Eye
(Typically 1.5K)
THE PROBLEM SCENARIO
INPUT OUTPUT
Right Eye
(Typically 6K)

3D & Motion Alignment Based Stereo Sequence
Processing t-2 t-1 t t+1 t+2
Left w o l f l f o w s t e r e o f l f l o w o w
Right w o l f l f o w s t e r e o f l f l o w o w
Left
• Highlights :
– Scintillation effect is reduced.
– Occlusion regions are better handled.
Right t-1 t t+1 t+2 t+3

SYNTHESIS RESULT ON REAL FOOTAGE

IMPLICATIONS FOR IMMERSIVE IBMR
CAMERA CONFIGURATIONS
Lo-res camera
Hi-res camera
Multi-resolution camera configuration allows 3D capture at the highest resolution as well as user-controlled large range of zooms without the need for zoom control on the cameras.

Original
Video
Site model
Georegistration of video to site model
Model-Centric Video Visualization
OR
Video-Centric Model Visualization
[Hsu,Supun,Kumar,Sawhney CVPR00]
Re-projection of video after merging with model.

Video to Site Model Alignment
• Model to frame alignment
REFINE
Correspondence-less exterior orientation from 3D-2D line pairs

0°
90°
Oriented Energy Pyramid
• Goal: representation which indicates edge strength in the image at various orientations and scales
• Orientation selectivity: reduce false matches
• Coarse-to-fine: increase capture range
45°
135°

Pose Refinement Algorithm
…iterative coarse to fine adjustment of pose ...
This will be an animation of the gradual improvement of alignment during the coarse to fine iterations regsite_animation.avi

Geo-Registration
Video to Reference Database Alignment
[Wildes et al. ICCV01]
Current Video 3D Reference Imagery

Registration : Radical Appearance Changes

Dynamic 3D Capture & Rendering
…global modeling is not feasible...
• Recovering depth from local views
• Depth refinement across multiple local views
• New view synthesis using multiple local views
Cross view depth checking

3D Shape/Depth Estimation from Multiple
Views of a Scene
Stereo Pair
• Estimation of high quality, artifact free depth maps coregistered with video imagery for rendering new views.
• Must work both outdoors and indoors

Multi-baseline depth estimation - requirements
[Tao,Sawhney,Kumar WACV00, ICCV01]
Accurate boundaries Accurate boundaries
Thin structures
Depth maps
A traditional stereo algorithm
New view rendering
Global matching method

New view rendering using local depth estimation
Local flow estimation
(1992)
Multiwindow plane+ parallax algorithm
(1998)
Color segmentation based stereo algorithm
(2000)
New view rendering

• Motivations
– be able to handle textureless regions
– handle object boundaries accurately
– global visibility constraints should be enforced
– Hypothesize reasonable depths for unmatched regions
• Solutions
– Global matching method - an analysis-by-synthesis approach
– Representation - smooth depth representation in homogeneous region
– Search method - neighborhood depth hypotheses generation
– Efficient algorithm - incremental warping
– Scene constraints - prior functions

Original image (frame 12) Original image (left)
Color segmentation [Comanicius 97]

New view rendering using local depth estimation
Left image True depth
Color segmentation based stereo algorithm new view rendering

Depth computation from 3 views
Video frame 11 Video frame 12 Video frame 13
Color segmentation (frame 12) Depth map (frame 12)

Multiple View Depth Recovery and New View
Rendering
New view rendering from a single view. left: from frame 212 , right: from frame 215
New view rendering from multiple views.

Multiple view depth recovery and new view rendering
Original 14 video frames (frame 04-17)
New view rendering (71 frames)
Depth map of frame 12 and 15

Immersive Visualization of a Dynamic Event
• Temporally consistent motion and 3D shape extraction
• Scintillation free dynamic high-quality rendering

AN IMMERSIVE IBMR GRAND CHALLENGE

AND IF WE DO IT RIGHT