multi-view stereo via volumetric graph-cuts george vogiatzis roberto cipolla cambridge univ....

35
Multi-view Stereo via Multi-view Stereo via Volumetric Graph-cuts Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing Oxford Brookes University

Upload: william-morefield

Post on 28-Mar-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Multi-view Stereo via Multi-view Stereo via Volumetric Graph-cutsVolumetric Graph-cuts

George Vogiatzis

Roberto Cipolla

Cambridge Univ. Engineering Dept.

Philip H. S. Torr

Department of ComputingOxford Brookes University

Page 2: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Multi-view Dense StereoMulti-view Dense Stereo

Calibrated images of Lambertian scene

3D model of scene

Page 3: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Multi-view Dense StereoMulti-view Dense Stereo

Volumetric

• Two main approaches• Volumetric• Disparity (depth) map

Page 4: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Dense Stereo Dense Stereo reconstruction problem:reconstruction problem:

Disparity-map

• Two main approaches• Volumetric• Disparity (depth) map

Page 5: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Shape representation Shape representation

• Disparity-maps• MRF formulation – good optimisation

techniques exist (Graph-cuts, Loopy BP)• MRF smoothness is viewpoint dependent• Disparity is unique per pixel – only

functions represented

Page 6: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Shape representation Shape representation

• Volumetric – e.g. Level-sets, Space carving etc.• Able to cope with non-functions• Levelsets: Local optimization• Space carving: no simple way to

impose surface smoothness

Page 7: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Our approachOur approach

• Cast volumetric methods in MRF framework

• Use approximate surface containing the real scene surface• E.g. visual hull

• Benefits:• General surfaces can be represented• No depth map merging required• Optimisation is tractable (MRF solvers)• Smoothness is viewpoint independent

Page 8: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Volumetric Graph cuts for Volumetric Graph cuts for segmentationsegmentation

• Volume is discretized • A binary MRF is defined on the voxels • Voxels are labelled as OBJECT and

BACKGROUND• Labelling cost set by OBJECT / BACKGROUND

intensity statistics• Compatibility cost set by intensity gradient

Boykov and Jolly ICCV 2001

Page 9: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Volumetric Graph cuts for Volumetric Graph cuts for stereostereo

Challenges:

• What do the two labels represent• How to define cost of setting them

• How to deal with occlusion• Interactions between distant voxels

Page 10: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Volumetric Graph cutsVolumetric Graph cuts

(x)

1. Outer surface

2. Inner surface (at constant offset)

3. Discretize middle volume

4. Assign photoconsistency cost to voxels

Page 11: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Volumetric Graph cutsVolumetric Graph cuts

Source

Sink

Page 12: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Volumetric Graph cutsVolumetric Graph cuts

Source

Sink

Cost of a cut (x)

dS

S

S

cut 3D Surface S

[Boykov and Kolmogorov ICCV 2001]

Page 13: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Volumetric Graph cutsVolumetric Graph cuts

Source

Sink

Minimum cut Minimal 3D Surface under photo-consistency metric

[Boykov and Kolmogorov ICCV 2001]

Page 14: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Photo-consistencyPhoto-consistency

• Occlusion

1. Get nearest point on outer surface

2. Use outer surface for occlusions

2. Discard occluded views

Page 15: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Photo-consistencyPhoto-consistency

• Occlusion

Self occlusion

Page 16: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Photo-consistencyPhoto-consistency

• Occlusion

Self occlusion

Page 17: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Photo-consistencyPhoto-consistency

• Occlusion

N

threshold on angle between normal and viewing direction

threshold= ~60

Page 18: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Photo-consistencyPhoto-consistency

• ScoreNormalised cross correlation

Use all remaining cameras pair wise

Average all NCC scores

Page 19: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Photo-consistencyPhoto-consistency

• ScoreAverage NCC = C

Voxel score

= 1 - exp( -tan2[(C-1)/4] / 2 )

0 1

= 0.05 in all experiments

Page 20: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

ExampleExample

Page 21: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Example - Visual Hull Example - Visual Hull

Page 22: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Example - SliceExample - Slice

Page 23: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Example - Slice with graphcutExample - Slice with graphcut

Page 24: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Example – 3DExample – 3D

Page 25: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Protrusion problemProtrusion problem

• ‘Balooning’ force• favouring bigger volumes that fill the visual hull

L.D. Cohen and I. Cohen. Finite-element methods for active contour models and balloons for 2-d and 3-d images. PAMI, 15(11):1131–1147, November 1993.

Page 26: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Protrusion problemProtrusion problem

• ‘Balooning’ force• favouring bigger volumes that fill the visual hull

L.D. Cohen and I. Cohen. Finite-element methods for active contour models and balloons for 2-d and 3-d images. PAMI, 15(11):1131–1147, November 1993.

(x) dS -

dV

S V

Page 27: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Protrusion problemProtrusion problem

Page 28: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

Protrusion problemProtrusion problem

Page 29: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

wij

SOURCE

wb

wb

GraphGraph

h

ji

wb = h3

wij = 4/3h2 * (i+j)/2[Boykov and Kolmogorov ICCV 2001]

Page 30: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

ResultsResults

• Model House

Page 31: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

ResultsResults

• Model House – Visual Hull

Page 32: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

ResultsResults

• Model House

Page 33: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

ResultsResults• Stone carving

Page 34: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

ResultsResults

• Haniwa

Page 35: Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing

SummarySummary

• Novel formulation for multiview stereo

• Volumetric scene representation

• Computationally tractable global optimisation using Graph-cuts.

• Visual hull for occlusions and geometric constraint

• Occlusions approximately modelled

Questions ?