Marc Levoy 1 Szymon Rusinkiewicz 1 Matt Ginzton 1 Jeremy Ginsberg 1 Kari Pulli 1 David Koller 1 Sean...
-
date post
20-Jan-2016 -
Category
Documents
-
view
220 -
download
0
Transcript of Marc Levoy 1 Szymon Rusinkiewicz 1 Matt Ginzton 1 Jeremy Ginsberg 1 Kari Pulli 1 David Koller 1 Sean...
Marc Levoy Marc Levoy 11
Szymon Szymon Rusinkiewicz Rusinkiewicz 11
Matt Ginzton Matt Ginzton 11
Jeremy Ginsberg Jeremy Ginsberg 11
Kari Pulli Kari Pulli 11
David Koller David Koller 11
Sean Anderson Sean Anderson 11
Jonathan Shade Jonathan Shade 22
Brian Curless Brian Curless 22
Lucas Pereira Lucas Pereira 11
James Davis James Davis 11
Duane Fulk Duane Fulk 33
11Computer Science Computer Science Department Stanford Department Stanford
UniversityUniversity22Department of Department of
Computer Science and Computer Science and Engineering University of Engineering University of
WashingtonWashington33Cyberware Inc.Cyberware Inc.
Presented Presented by : Abhinav by : Abhinav DayalDayal
Overview Overview
MotivationsMotivations• push 3D scanning technology push 3D scanning technology • tool for art historianstool for art historians• lasting archivelasting archive
Technical goalsTechnical goals• scan a big statue scan a big statue • capture chisel markscapture chisel marks• capture reflectancecapture reflectance
5 meters
1/4 mm1/4 mm
20,000:1
20,0002
1 billion
Why capture chisel Why capture chisel marks?marks?
Atlas (Accademia)
ugnetto
?
Day (Medici Chapel)
2 mm
single scan of St. Matthew 1 mm
Issues InvolvedIssues Involved
ScanningScanning Scanner designScanner design Scanning ProcedureScanning Procedure
Post ProcessingPost Processing Range ProcessingRange Processing Color ProcessingColor Processing
Handling large data setsHandling large data sets
Scanner DesignScanner Design
Laser stripe triangulation systemLaser stripe triangulation system Resolution and field of view Resolution and field of view
Should capture chisel marksShould capture chisel marks Reasonable size of resulting datasetReasonable size of resulting dataset
Standoff and baselineStandoff and baseline Longer standoff Longer standoff access to deeper access to deeper
recesses + safe distance from statuerecesses + safe distance from statue Longer standoff Longer standoff longer baselinelonger baseline
prone to miscalibrationprone to miscalibration
Color AcquisitionColor Acquisition
Single passSingle pass 1D luminaire and 1D color sensor1D luminaire and 1D color sensor
Cross talk b/w luminaire & sensorCross talk b/w luminaire & sensor poor fidelity poor fidelity RGB lasersRGB lasers
Large and complexLarge and complex
Separate passSeparate pass They used broadband luminaire and They used broadband luminaire and
separate sensor (Digital Camera – 1520 x separate sensor (Digital Camera – 1520 x 1144 pix res)1144 pix res)
Camera Resolution and Camera Resolution and fieldfield
For color to range calibration For color to range calibration match respective standoffsmatch respective standoffs
One color per pixel resolution One color per pixel resolution (per (per range data)range data)
Controlled illuminationControlled illumination Depth of fieldDepth of field
DOF > FOV(Z) of range-cameraDOF > FOV(Z) of range-camera
Gantry:Geometric Gantry:Geometric designdesign
4 motorized axes
laser, range camera,white light, and color camera
truss extensionsfor tall statues
Working volume of the Working volume of the scannerscanner
calibrated motionscalibrated motions pitch pitch (yellow)(yellow) pan pan (blue)(blue) horizontal translation horizontal translation (orange)(orange)
• uncalibrated motions– vertical translation– remounting the scan head– moving the entire gantry
Scanning the DavidScanning the David
height of gantry: 7.5 height of gantry: 7.5 metersmeters
weight of gantry: 800 weight of gantry: 800 kilogramskilograms
CalibrationCalibrationRange CalibrationRange Calibration
Color CalibrationColor Calibration
Their calibration was complex and did moderately Their calibration was complex and did moderately wellwell
• Used a planar target with feature points to Used a planar target with feature points to calculate camera’s intrinsic parameters and calculate camera’s intrinsic parameters and to build a per pixel intensity correction to build a per pixel intensity correction tabletable
Scanning procedureScanning procedure
Range ScanningRange Scanning Typical range involves several Typical range involves several
horizontally adjacent vertical sweeps or horizontally adjacent vertical sweeps or vice versavice versa
Overlap adjacent sweeps by 40% Overlap adjacent sweeps by 40% Overlap adjacent shells by 15% Overlap adjacent shells by 15%
Color ScanningColor Scanning Image with spotlight – Image without Image with spotlight – Image without
spotlight = Image with only spotlightspotlight = Image with only spotlight
Safety ConcernsSafety Concerns
energy depositionenergy deposition Low and not a problemLow and not a problem
avoiding collisionsavoiding collisions manual motion controlsmanual motion controls automatic cutoff switchesautomatic cutoff switches one person serves as spotterone person serves as spotter
surviving collisionssurviving collisions pad the scan headpad the scan head
Range processing Range processing pipelinepipeline stepssteps
manual initial alignmentmanual initial alignment ICP to one existing scanICP to one existing scan automatic ICP of all automatic ICP of all
overlapping pairsoverlapping pairs global relaxation to spread out global relaxation to spread out
errorerror merging using volumetric merging using volumetric
methodmethod(space carving)(space carving)
problemsproblems should have tracked the gantry should have tracked the gantry
locationlocation ICP is unstable on smooth ICP is unstable on smooth
surfacessurfaces
Color processing Color processing pipelinepipeline
stepssteps compensate for ambient compensate for ambient
illuminationillumination discard shadowed or specular pixelsdiscard shadowed or specular pixels map onto vertices – one color per map onto vertices – one color per
vertexvertex correct for irradiance correct for irradiance diffuse diffuse
reflectancereflectance
limitationslimitations ignored interreflectionsignored interreflections ignored subsurface scatteringignored subsurface scattering treated diffuse as Lambertiantreated diffuse as Lambertian used aggregate surface normalsused aggregate surface normals
Handling large Handling large datasetsdatasets
range images instead of polygon range images instead of polygon meshesmeshes r(u,v) (special case of displacement r(u,v) (special case of displacement
map) yields 18:1 lossless map) yields 18:1 lossless compression (run-length encoding)compression (run-length encoding)
multiresolution using (range) multiresolution using (range) image pyramidimage pyramid
lazy evaluationlazy evaluation viewer based on point rendering viewer based on point rendering
(Qsplat)(Qsplat)
Range image Range image pyramidspyramids
Range image at 1x resolutionRange image at 1x resolutionRange image at 1/2x resolutionRange image at 1/2x resolution
Range image at 1/4x resolutionRange image at 1/4x resolution
Samples with any of Samples with any of its four children at the its four children at the next finer pyramid next finer pyramid level missing deletedlevel missing deleted
Redden parents of Redden parents of missing children in missing children in proportion to the proportion to the fraction of its missing fraction of its missing childrenchildren
Can see holes at Can see holes at coarser coarser resolutionsresolutions
Problems facedProblems faced
Approximated marble as truly lambertianApproximated marble as truly lambertian Many unavoidable holes in the scanMany unavoidable holes in the scan Expensive and bulky gantryExpensive and bulky gantry Inadequate calibrationInadequate calibration Manual view planning Manual view planning prone to errors prone to errors Manual alignment of successive scansManual alignment of successive scans
How optically cooperative is How optically cooperative is marble?marble?
• systematic bias of 40 micronssystematic bias of 40 microns
• noise of 150 – 250 micronsnoise of 150 – 250 microns
– worse at oblique angles of incidenceworse at oblique angles of incidence
– worse for polished statuesworse for polished statues
Removing the Removing the holesholes Space carving techniqueSpace carving technique
Create a maximum region of space Create a maximum region of space consistent with the scansconsistent with the scans
Creates a watertight surface Creates a watertight surface may lead to surfaces that are less may lead to surfaces that are less
plausible than smoothly extending plausible than smoothly extending the observed surfacesthe observed surfaces.
Recent work based upon Recent work based upon Volumetric diffusion (Volumetric diffusion (Video))
Reference: Filling holes in complex surfaces using volumetric diffusion: James Davis, Steven R. Marschner, Matt Garr, Marc Levoy, Computer Science Department, Stanford University (August, 2001)
ConclusionConclusion provides remarkable background for provides remarkable background for
future developmentsfuture developments Provides massive mesh modelsProvides massive mesh models Discusses basic issues involved in 3D Discusses basic issues involved in 3D
scanningscanning Many lessons learntMany lessons learnt Encouraging Encouraging many new applicationsmany new applications ((in areas in areas
such as: reverse engineering; industrial design; repair, such as: reverse engineering; industrial design; repair, reproduction, and improvement of machinery; medical reproduction, and improvement of machinery; medical diagnostics, analysis and simulation; 3D photography; diagnostics, analysis and simulation; 3D photography; and building rich virtual environmentsand building rich virtual environments))
This is just a beginningThis is just a beginning
QUESTIONS?QUESTIONS?