Multi-view Stereo via Volumetric Graph-cuts George Vogiatzis Roberto Cipolla Cambridge Univ. Engineering Dept. Philip H. S. Torr Department of Computing Oxford Brookes University
Multi-view Dense Stereo Calibrated images of Lambertian scene 3D model of scene
Multi-view Dense Stereo Two main approaches Volumetric Disparity (depth) map Volumetric
Dense Stereo reconstruction problem: Two main approaches Volumetric Disparity (depth) map Disparity-map
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
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
Our 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
Volumetric Graph cuts for segmentation Boykov and Jolly ICCV 2001 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
Volumetric Graph cuts for stereo Challenges: What do the two labels represent How to define cost of setting them How to deal with occlusion Interactions between distant voxels
Volumetric Graph cuts (x) 1. Outer surface 2. Inner surface (at constant offset) (x) 3. Discretize middle volume 4. Assign photoconsistency cost to voxels
Volumetric Graph cuts Source Sink
Volumetric Graph cuts S cut 3D Surface S Cost of a cut (x) dS Source [Boykov and Kolmogorov ICCV 2001] S S Sink
Volumetric Graph cuts Minimum cut Minimal 3D Surface under photo-consistency metric Source [Boykov and Kolmogorov ICCV 2001] Sink
Photo-consistency Occlusion 1. Get nearest point on outer surface 2. Use outer surface for occlusions 2. Discard occluded views
Photo-consistency Occlusion Self occlusion
Photo-consistency Occlusion Self occlusion
Photo-consistency Occlusion threshold on angle between normal and viewing direction threshold= ~60 N
Photo-consistency Score Normalised cross correlation Use all remaining cameras pair wise Average all NCC scores Score
Photo-consistency Score = 1 - exp( -tan2[(C-1)/4] / 2 ) Average NCC = C Voxel score = 1 - exp( -tan2[(C-1)/4] / 2 ) Score 0 1 = 0.05 in all experiments
Example
Example - Visual Hull
Example - Slice
Example - Slice with graphcut
Example – 3D
Protrusion 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 Protrusion 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.
Protrusion problem
Protrusion problem
Graph wij = 4/3h2 * (i+j)/2 wb wb = h3 wij i j h SOURCE [Boykov and Kolmogorov ICCV 2001] wb = h3 wij i j h
Results Model House
Results Model House – Visual Hull
Results Model House
Results Stone carving
Results Haniwa
Summary Questions ? 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 ?