Hoip10 presentación reconstrucción de superficies_upc

62

description

Presentación de la Universidad Politécnica de Catalunya sobre reconstrucciónde superficies realizada durante las jornadas HOIP 2010 organizadas por la Unidad de Sistemas de Información e Interacción TECNALIA. Más información en http://www.tecnalia.com/es/ict-european-software-institute/index.htm

Transcript of Hoip10 presentación reconstrucción de superficies_upc

Page 1: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Surface Reconstructionby

Restricted and Oriented Propagation

Xavier Suau Josep R. Casas Javier Ruiz-Hidalgo

{xavier.suau, josep.ramon.casas, j.ruiz}@upc.edu

Universitat Politècnica de Catalunya

November 16, 2010

Page 2: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Outline

1 Motivation and state of the art

2 Propagation Algorithm

3 Experimental Results

4 Conclusion

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 1 / 20

Page 3: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Context

Large 3D point clouds are very common datasets, being mostly obtained from:

Laser scans Multiview datasets Virtual datasets

The objective is to have a meshed representation of these type of datasets

in this case, for visualization purposes

in a fast, up to real-time, way

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 2 / 20

Page 4: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

State of the Art

Results are evaluated against a reference composed of:

Ball-Pivoting Algorithm

• Very accuratereconstruction

• Sensitive to densityvariations

Poisson Reconstruction

• Watertight reconstructedsurface

• Fast reconstructions providelow level of detail

Marching Cubes + APSS

• Watertight reconstructedsurface

• Voxelization required

all of them implemented in the MeshLab c© software

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 3 / 20

Page 5: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

State of the Art

Results are evaluated against a reference composed of:

Ball-Pivoting Algorithm

• Very accuratereconstruction

• Sensitive to densityvariations

Poisson Reconstruction

• Watertight reconstructedsurface

• Fast reconstructions providelow level of detail

Marching Cubes + APSS

• Watertight reconstructedsurface

• Voxelization required

all of them implemented in the MeshLab c© software

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 3 / 20

Page 6: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

State of the Art

Results are evaluated against a reference composed of:

Ball-Pivoting Algorithm

• Very accuratereconstruction

• Sensitive to densityvariations

Poisson Reconstruction

• Watertight reconstructedsurface

• Fast reconstructions providelow level of detail

Marching Cubes + APSS

• Watertight reconstructedsurface

• Voxelization required

all of them implemented in the MeshLab c© software

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 3 / 20

Page 7: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

State of the Art

Results are evaluated against a reference composed of:

Ball-Pivoting Algorithm

• Very accuratereconstruction

• Sensitive to densityvariations

Poisson Reconstruction

• Watertight reconstructedsurface

• Fast reconstructions providelow level of detail

Marching Cubes + APSS

• Watertight reconstructedsurface

• Voxelization required

all of them implemented in the MeshLab c© software

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 3 / 20

Page 8: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Outline

1 Motivation and state of the art

2 Propagation Algorithm

3 Experimental Results

4 Conclusion

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 4 / 20

Page 9: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Algorithm overview

From 3D point clouds...

...to meshed surfaces

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 5 / 20

Page 10: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Algorithm overview

From 3D point clouds...

...to meshed surfaces

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 5 / 20

Page 11: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Algorithm overview

From 3D point clouds...

...to meshed surfaces

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 5 / 20

Page 12: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Voxelization

• The target point cloud S is composed of points pi = (Pi , Ci ) withPi = (xi , yi , zi ) and Ci = (ri , gi , bi )

• Voxels υk are associated to pi as follows

0 points in voxel

υk ← ∅

1 point p = (P, C) in voxel

υk ← (P, C)

m points pj

υk ← (P, C)

Voxels υk 6= ∅ are called seed voxels, or υS

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 6 / 20

Page 13: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Voxelization

• The target point cloud S is composed of points pi = (Pi , Ci ) withPi = (xi , yi , zi ) and Ci = (ri , gi , bi )

• Voxels υk are associated to pi as follows

0 points in voxel

υk ← ∅

1 point p = (P, C) in voxel

υk ← (P, C)

m points pj

υk ← (P, C)

Voxels υk 6= ∅ are called seed voxels, or υS

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 6 / 20

Page 14: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Pattern

Propagation, why? To �nd close neighbors in the discretized space

How? With a propagation pattern or set of positions relative to a seed voxel

Omni-26 Omni-18 Omni-6 6DO Oriented Pattern

Knowing that direction of neighbor �nding is indi�erent

Omni patterns check both directions, redundant!

The 6DO Oriented Pattern

• Reduces the amount of redundant edges

• Is faster than Omni-18 with the same spatial coverage

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 7 / 20

Page 15: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Pattern

Propagation, why? To �nd close neighbors in the discretized space

How? With a propagation pattern or set of positions relative to a seed voxel

Omni-26 Omni-18 Omni-6 6DO Oriented Pattern

Knowing that direction of neighbor �nding is indi�erent

Omni patterns check both directions, redundant!

The 6DO Oriented Pattern

• Reduces the amount of redundant edges

• Is faster than Omni-18 with the same spatial coverage

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 7 / 20

Page 16: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Pattern

Propagation, why? To �nd close neighbors in the discretized space

How? With a propagation pattern or set of positions relative to a seed voxel

Omni-26 Omni-18 Omni-6 6DO Oriented Pattern

Knowing that direction of neighbor �nding is indi�erent

Omni patterns check both directions, redundant!

The 6DO Oriented Pattern

• Reduces the amount of redundant edges

• Is faster than Omni-18 with the same spatial coverage

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 7 / 20

Page 17: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Pattern

Propagation, why? To �nd close neighbors in the discretized space

How? With a propagation pattern or set of positions relative to a seed voxel

Omni-26 Omni-18 Omni-6 6DO Oriented Pattern

Knowing that direction of neighbor �nding is indi�erent

Omni patterns check both directions, redundant!

The 6DO Oriented Pattern

• Reduces the amount of redundant edges

• Is faster than Omni-18 with the same spatial coverage

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 7 / 20

Page 18: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 19: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 20: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 21: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 22: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 23: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 24: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 25: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Propagation Steps

Iterative Algorithm

• Propagation starts at every seed voxel υSi

• Voxels ∈ 6DO are associated to its seedvoxels υSi , building up seed volumes Vithat grow at every iteration

• At propagation end, intersections Vi ∩ Vjde�ne pairs of neighbors pi ! pj

• Triangular faces are obtained from the listof neighbors

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 8 / 20

Page 26: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Stop Threshold

Propagation iterations should be stopped at the appropriate moment to avoidmeshing distant points

Edge Density

• The number of created edges per iteration is called edge density or De

• De presents a �rst maximumDemax at a low number ofiterations κmax, which correspondsto the meshing of the main surface

• Propagation stops at iteration k

which veri�es:(κ ≥ 2κmax)

)∧(De(κ) <

1

4Demax

)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 9 / 20

Page 27: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Stop Threshold

Propagation iterations should be stopped at the appropriate moment to avoidmeshing distant points

Edge Density

• The number of created edges per iteration is called edge density or De

• De presents a �rst maximumDemax at a low number ofiterations κmax, which correspondsto the meshing of the main surface

• Propagation stops at iteration k

which veri�es:(κ ≥ 2κmax)

)∧(De(κ) <

1

4Demax

)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 9 / 20

Page 28: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Stop Threshold

Propagation iterations should be stopped at the appropriate moment to avoidmeshing distant points

Edge Density

• The number of created edges per iteration is called edge density or De

• De presents a �rst maximumDemax at a low number ofiterations κmax, which correspondsto the meshing of the main surface

• Propagation stops at iteration k

which veri�es:(κ ≥ 2κmax)

)∧(De(κ) <

1

4Demax

)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 9 / 20

Page 29: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Stop Threshold

Propagation iterations should be stopped at the appropriate moment to avoidmeshing distant points

Edge Density

• The number of created edges per iteration is called edge density or De

• De presents a �rst maximumDemax at a low number ofiterations κmax, which correspondsto the meshing of the main surface

• Propagation stops at iteration k

which veri�es:(κ ≥ 2κmax)

)∧(De(κ) <

1

4Demax

)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 9 / 20

Page 30: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Outline

1 Motivation and state of the art

2 Propagation Algorithm

3 Experimental Results

4 Conclusion

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 10 / 20

Page 31: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation of results

Quantitative Evaluation

Two main characteristics are evaluated:

δH Hausdor� Distance metric between a groundtruth surface and areconstructed surface

tO Overall calculation time on a 64-bit Intel Xeon CPU @ 3.00GHz processor(includes memory allocation and mesh writing)

Results are presented on an Accuracy Vs. Speed (δH , tO) plane

Qualitative Evaluation

Global visual inspection

Four 3D models provided by the Stanford 3D Scanning Repository are tested:

Bunny Hand Dragon Happy Buddha

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 11 / 20

Page 32: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation of results

Quantitative Evaluation

Two main characteristics are evaluated:

δH Hausdor� Distance metric between a groundtruth surface and areconstructed surface

tO Overall calculation time on a 64-bit Intel Xeon CPU @ 3.00GHz processor(includes memory allocation and mesh writing)

Results are presented on an Accuracy Vs. Speed (δH , tO) plane

Qualitative Evaluation

Global visual inspection

Four 3D models provided by the Stanford 3D Scanning Repository are tested:

Bunny Hand Dragon Happy Buddha

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 11 / 20

Page 33: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation of results

Quantitative Evaluation

Two main characteristics are evaluated:

δH Hausdor� Distance metric between a groundtruth surface and areconstructed surface

tO Overall calculation time on a 64-bit Intel Xeon CPU @ 3.00GHz processor(includes memory allocation and mesh writing)

Results are presented on an Accuracy Vs. Speed (δH , tO) plane

Qualitative Evaluation

Global visual inspection

Four 3D models provided by the Stanford 3D Scanning Repository are tested:

Bunny Hand Dragon Happy Buddha

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 11 / 20

Page 34: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation of results

Quantitative Evaluation

Two main characteristics are evaluated:

δH Hausdor� Distance metric between a groundtruth surface and areconstructed surface

tO Overall calculation time on a 64-bit Intel Xeon CPU @ 3.00GHz processor(includes memory allocation and mesh writing)

Results are presented on an Accuracy Vs. Speed (δH , tO) plane

Qualitative Evaluation

Global visual inspection

Four 3D models provided by the Stanford 3D Scanning Repository are tested:

Bunny Hand Dragon Happy Buddha

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 11 / 20

Page 35: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation of results

Quantitative Evaluation

Two main characteristics are evaluated:

δH Hausdor� Distance metric between a groundtruth surface and areconstructed surface

tO Overall calculation time on a 64-bit Intel Xeon CPU @ 3.00GHz processor(includes memory allocation and mesh writing)

Results are presented on an Accuracy Vs. Speed (δH , tO) plane

Qualitative Evaluation

Global visual inspection

Four 3D models provided by the Stanford 3D Scanning Repository are tested:

Bunny Hand Dragon Happy Buddha

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 11 / 20

Page 36: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Voxelization e�ect

Voxelization resolution is ReOP's critical parameter

• Low resolution: Poor visual quality

• High resolution: Higher calculation time and memory requirements

76×57×34 voxels11,145 vertices76,124 faces

1.2 s

226×170×101 voxels85,082 vertices529,916 faces

8.9 s

376×283×168 voxels181,509 vertices994,578 faces

17.3 s

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 12 / 20

Page 37: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Happy Buddha dataset (543,652 points)

(δH , tO) plane Point Cloud

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 13 / 20

Page 38: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Happy Buddha dataset (543,652 points)

(δH , tO) plane Ball-Pivoting

238, 193 faces

(δH , tO ) = (0.000719, 1429 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 13 / 20

Page 39: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Happy Buddha dataset (543,652 points)

(δH , tO) plane MCubes+APSS

2, 641, 481 faces

(δH , tO ) = (0.000046, 528 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 13 / 20

Page 40: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Happy Buddha dataset (543,652 points)

(δH , tO) plane Poisson Reconstruction

631, 480 faces

(δH , tO ) = (0.000184, 65.1 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 13 / 20

Page 41: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Happy Buddha dataset (543,652 points)

(δH , tO) plane ReOP

1, 367, 336 faces

(δH , tO ) = (0.000031, 22.2 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 13 / 20

Page 42: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Comparative (Happy Buddha - 543,652 points)

Ball-Pivoting

238, 193 faces

(δH , tO ) =

(0.000719, 1429 s)

MCubes+APSS

2, 641, 481 faces

(δH , tO ) =

(0.000046, 528 s)

Poisson Rec.

631, 480 faces

(δH , tO ) =

(0.000184, 65.1 s)

ReOP

1, 367, 336 faces

(δH , tO ) =

(0.000031, 22.2 s)

Results on Happy Buddha, largest dataset

• About 23x faster than MCubes+APSS for a similar good quality

• Reasonable amount of faces, about 2.5 · Npoints

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 14 / 20

Page 43: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Comparative (Happy Buddha - 543,652 points)

Ball-Pivoting

238, 193 faces

(δH , tO ) =

(0.000719, 1429 s)

MCubes+APSS

2, 641, 481 faces

(δH , tO ) =

(0.000046, 528 s)

Poisson Rec.

631, 480 faces

(δH , tO ) =

(0.000184, 65.1 s)

ReOP

1, 367, 336 faces

(δH , tO ) =

(0.000031, 22.2 s)

Results on Happy Buddha, largest dataset

• About 23x faster than MCubes+APSS for a similar good quality

• Reasonable amount of faces, about 2.5 · Npoints

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 14 / 20

Page 44: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Comparative (Happy Buddha - 543,652 points)

Ball-Pivoting

238, 193 faces

(δH , tO ) =

(0.000719, 1429 s)

MCubes+APSS

2, 641, 481 faces

(δH , tO ) =

(0.000046, 528 s)

Poisson Rec.

631, 480 faces

(δH , tO ) =

(0.000184, 65.1 s)

ReOP

1, 367, 336 faces

(δH , tO ) =

(0.000031, 22.2 s)

Results on Happy Buddha, largest dataset

• About 23x faster than MCubes+APSS for a similar good quality

• Reasonable amount of faces, about 2.5 · Npoints

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 14 / 20

Page 45: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Stanford Bunny dataset (35,947 points)

(δH , tO) plane Point Cloud

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 15 / 20

Page 46: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Stanford Bunny dataset (35,947 points)

(δH , tO) plane Ball-Pivoting

238, 193 faces

(δH , tO ) = (0.000113, 8.2 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 15 / 20

Page 47: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Stanford Bunny dataset (35,947 points)

(δH , tO) plane MCubes+APSS

2, 641, 481 faces

(δH , tO ) = (0.000042, 23 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 15 / 20

Page 48: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Stanford Bunny dataset (35,947 points)

(δH , tO) plane Poisson Reconstruction

631, 480 faces

(δH , tO ) = (0.000285, 10.3 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 15 / 20

Page 49: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Evaluation on the (δH , tO) plane

Stanford Bunny dataset (35,947 points)

(δH , tO) plane ReOP

1, 367, 336 faces

(δH , tO ) = (0.000044, 0.96 s)

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 15 / 20

Page 50: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Comparative (Stanford Bunny - 35,947 points)

Ball-Pivoting

70, 832 faces

(δH , tO ) =

(0.000113, 8.2 s)

MCubes+APSS

769, 029 faces

(δH , tO ) =

(0.000042, 23 s)

Poisson Rec.

70, 438 faces

(δH , tO ) =

(0.000285, 10.3 s)

ReOP

147, 029 faces

(δH , tO ) =

(0.000044, 0.96 s)

Results on Stanford Bunny, smallest dataset

• About 23x faster than MCubes+APSS for a the same quality

• Reasonable amount of faces, about 3 · Npoints

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 16 / 20

Page 51: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Comparative (Stanford Bunny - 35,947 points)

Ball-Pivoting

70, 832 faces

(δH , tO ) =

(0.000113, 8.2 s)

MCubes+APSS

769, 029 faces

(δH , tO ) =

(0.000042, 23 s)

Poisson Rec.

70, 438 faces

(δH , tO ) =

(0.000285, 10.3 s)

ReOP

147, 029 faces

(δH , tO ) =

(0.000044, 0.96 s)

Results on Stanford Bunny, smallest dataset

• About 23x faster than MCubes+APSS for a the same quality

• Reasonable amount of faces, about 3 · Npoints

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 16 / 20

Page 52: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Comparative (Stanford Bunny - 35,947 points)

Ball-Pivoting

70, 832 faces

(δH , tO ) =

(0.000113, 8.2 s)

MCubes+APSS

769, 029 faces

(δH , tO ) =

(0.000042, 23 s)

Poisson Rec.

70, 438 faces

(δH , tO ) =

(0.000285, 10.3 s)

ReOP

147, 029 faces

(δH , tO ) =

(0.000044, 0.96 s)

Results on Stanford Bunny, smallest dataset

• About 23x faster than MCubes+APSS for a the same quality

• Reasonable amount of faces, about 3 · Npoints

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 16 / 20

Page 53: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Outline

1 Motivation and state of the art

2 Propagation Algorithm

3 Experimental Results

4 Conclusion

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 17 / 20

Page 54: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

The presented ReOP algorithm is...

• Surface reconstruction is performed about 23x faster than the reference,for a given quality

• ReOP quality is similar to the best reference method

• ReOP reconstructed mesh is visually clear and presents few artifacts

• The seed voxel/volume structure is suitable to be parallelized on GPU

• The output mesh has no manifold properties

ReOP is suitable for...

• Real-time applications with small datasets (<50,000 points inexperiments)

• Large datasets reconstruction (millions of points), such those obtained inmultiview applications

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 18 / 20

Page 55: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

The presented ReOP algorithm is...

• Surface reconstruction is performed about 23x faster than the reference,for a given quality

• ReOP quality is similar to the best reference method

• ReOP reconstructed mesh is visually clear and presents few artifacts

• The seed voxel/volume structure is suitable to be parallelized on GPU

• The output mesh has no manifold properties

ReOP is suitable for...

• Real-time applications with small datasets (<50,000 points inexperiments)

• Large datasets reconstruction (millions of points), such those obtained inmultiview applications

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 18 / 20

Page 56: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

The presented ReOP algorithm is...

• Surface reconstruction is performed about 23x faster than the reference,for a given quality

• ReOP quality is similar to the best reference method

• ReOP reconstructed mesh is visually clear and presents few artifacts

• The seed voxel/volume structure is suitable to be parallelized on GPU

• The output mesh has no manifold properties

ReOP is suitable for...

• Real-time applications with small datasets (<50,000 points inexperiments)

• Large datasets reconstruction (millions of points), such those obtained inmultiview applications

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 18 / 20

Page 57: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

The presented ReOP algorithm is...

• Surface reconstruction is performed about 23x faster than the reference,for a given quality

• ReOP quality is similar to the best reference method

• ReOP reconstructed mesh is visually clear and presents few artifacts

• The seed voxel/volume structure is suitable to be parallelized on GPU

• The output mesh has no manifold properties

ReOP is suitable for...

• Real-time applications with small datasets (<50,000 points inexperiments)

• Large datasets reconstruction (millions of points), such those obtained inmultiview applications

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 18 / 20

Page 58: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

The presented ReOP algorithm is...

• Surface reconstruction is performed about 23x faster than the reference,for a given quality

• ReOP quality is similar to the best reference method

• ReOP reconstructed mesh is visually clear and presents few artifacts

• The seed voxel/volume structure is suitable to be parallelized on GPU

• The output mesh has no manifold properties

ReOP is suitable for...

• Real-time applications with small datasets (<50,000 points inexperiments)

• Large datasets reconstruction (millions of points), such those obtained inmultiview applications

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 18 / 20

Page 59: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

The presented ReOP algorithm is...

• Surface reconstruction is performed about 23x faster than the reference,for a given quality

• ReOP quality is similar to the best reference method

• ReOP reconstructed mesh is visually clear and presents few artifacts

• The seed voxel/volume structure is suitable to be parallelized on GPU

• The output mesh has no manifold properties

ReOP is suitable for...

• Real-time applications with small datasets (<50,000 points inexperiments)

• Large datasets reconstruction (millions of points), such those obtained inmultiview applications

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 18 / 20

Page 60: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

The presented ReOP algorithm is...

• Surface reconstruction is performed about 23x faster than the reference,for a given quality

• ReOP quality is similar to the best reference method

• ReOP reconstructed mesh is visually clear and presents few artifacts

• The seed voxel/volume structure is suitable to be parallelized on GPU

• The output mesh has no manifold properties

ReOP is suitable for...

• Real-time applications with small datasets (<50,000 points inexperiments)

• Large datasets reconstruction (millions of points), such those obtained inmultiview applications

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 18 / 20

Page 61: Hoip10 presentación reconstrucción de superficies_upc

Motivation and SoAPropagation AlgorithmExperimental Results

Conclusion

Future work

• Adapt propagation pattern to topology and sampling density of surfaces

• Find faster structures for close neighbor queries (eg. kdtree)

• Obtain manifold meshes while preserving execution speed

• GPU implementation

Xavier Suau, Josep R. Casas, Javier Ruiz-Hidalgo Surface Reconstruction by ReOP 19 / 20

Page 62: Hoip10 presentación reconstrucción de superficies_upc

Thank You

Questions