Supplementary Material: Pixelwise View Selection for ...
3
Supplementary Material: Pixelwise View Selection for Unstructured Multi-View Stereo Johannes L. Sch¨ onberger 1 , Enliang Zheng 2 , Marc Pollefeys 1,3 , Jan-Michael Frahm 2 1 ETH Z¨ urich, 2 UNC Chapel Hill, 3 Microsoft {jsch,pomarc}@inf.ethz.ch, {ezheng,jmf}@cs.unc.edu 1 Algorithm The algorithm for computing depths and normals according to Section 4.6. Function: Coordinate Descent Input: Images X Output: Depths θ, Normals N For m = 1 to M Set X ref = X m Set X src = X \{X m } For i = 1 to I1 Sweep with Eq. (6) For i = 1 to I2 For m = 1 to M Set X ref = X m Set X src = X \{X m } Sweep with Eq. (11) Function: Sweep Input: Images X ref , X src Output: Depths θ ref , Normals N ref For r = 1 to 4 Eq. Rotate reference image by 90 ◦ For l = L to 1 For m = 1 to M Compute backward message ← - m m l (14) For l = 1 to L For m = 1 to M Compute forward message - → m m l (13) Compute q(Z m l ) (12) PatchMatch propagation/sampling (7) Estimate (θ * l , n * l ) over q l (θ l , n l ) (6/11) For m = 1 to M Recompute forward message - → m m l (13)
Transcript of Supplementary Material: Pixelwise View Selection for ...
Supplementary Material:Pixelwise View Selection for Unstructured
Multi-View Stereo
Johannes L. Schonberger1, Enliang Zheng2,Marc Pollefeys1,3, Jan-Michael Frahm2
1ETH Zurich, 2UNC Chapel Hill, 3Microsoft{jsch,pomarc}@inf.ethz.ch, {ezheng,jmf}@cs.unc.edu
1 Algorithm
The algorithm for computing depths and normals according to Section 4.6.
Function: Coordinate DescentInput: Images X
Output: Depths θ, Normals N
For m = 1 to M
Set Xref = Xm
Set Xsrc = X \ {Xm}For i = 1 to I1Sweep with Eq. (6)
For i = 1 to I2For m = 1 to M
Set Xref = Xm
Set Xsrc = X \ {Xm}Sweep with Eq. (11)
Function: Sweep
Input: Images Xref,Xsrc
Output: Depths θref, Normals Nref
For r = 1 to 4 Eq.Rotate reference image by 90◦
For l = L to 1For m = 1 to MCompute backward message ←−mm
l (14)For l = 1 to LFor m = 1 to MCompute forward message −→mm
l (13)Compute q(Zm
l ) (12)PatchMatch propagation/sampling (7)Estimate (θ∗l ,n
∗l ) over ql(θl,nl) (6/11)
For m = 1 to MRecompute forward message −→mm
l (13)
2 J.L. Schonberger, E. Zheng, M. Pollefeys, J.M. Frahm
2 Comparison
This section compares our results against Goesele et al. [43] on Internet photocollections [5]. From left to right: reference image, depth map by Goesele et al.,our depth and normal maps. From top to bottom: Big Ben, England; SagradaFamilia, Spain; Louvre, France; Florence Cathedral, Italy.
Goesele et al. Ours
Pixelwise View Selection for Unstructured Multi-View Stereo 3
3 Point Clouds
Examples of fused point clouds for reconstructed Internet photo datasets pro-duced by Heinly et al. [5]. From top left to bottom right: Milan Cathedral, Italy;Piazza dei Miracoli, Italy; Reichstag, Germany; Temple in Kyoto, Japan; St. Vi-tus Cathedral, Czech Republic; Piazza San Marco, Italy; St. Paul’s Cathedral,England; Peter’s Dome, Vatican; Pantheon, Italy; Sagrada Familia (front), Spain;Sagrada Familia (back), Spain; British Museum, England; Florence Cathedral,Italy; Sistine Chapel, Vatican; Piazza della Signora, Italy; Piazza Public, Italy.
