BSP-Net: Generating Compact Meshes via Binary Space Partitioning - - PowerPoint PPT Presentation

▶
bsp net generating compact meshes via binary space
SMART_READER_LITE
LIVE PREVIEW

BSP-Net: Generating Compact Meshes via Binary Space Partitioning - - PowerPoint PPT Presentation

Feb 26, 2024 212 likes •547 views

BSP-Net: Generating Compact Meshes via Binary Space Partitioning Zhiqin Chen Andrea Tagliasacchi Hao (Richard) Zhang Simon Fraser University Google Research Simon Fraser University Presented by Zhiqin

slide-1
SLIDE 1

Presented by

陈之钦，西蒙弗雷泽大学

Zhiqin Chen, Ph.D. student, Simon Fraser University

BSP-Net: Generating Compact Meshes via Binary Space Partitioning

Zhiqin Chen Simon Fraser University Andrea Tagliasacchi Google Research Hao (Richard) Zhang Simon Fraser University

slide-2
SLIDE 2

3D shape representations

2 Mesh ShapeNet

Chang et al, arXiv

PolyGen

Nash et al, arXiv

slide-3
SLIDE 3

3D shape representations

3 Voxel 3D-R2N2

Choy et al, ECCV 2016

3DGAN

Wu et al, NIPS 2016

slide-4
SLIDE 4

3D shape representations

4 Point cloud

Learning Representations and Generative Models for 3D Point Clouds Achlioptas et al, ICML 2018

z (128d) 512 1024 6144 = 2048x3 FC FC FC

slide-5
SLIDE 5

3D shape representations

5 3DN

Wang et al, CVPR 2019

AtlasNet

Groueix et al, CVPR 2018

Warping template meshes Pixel2mesh

Wang et al, ECCV 2018

slide-6
SLIDE 6

3D shape representations

6 implicit field

IM-NET

Chen & Zhang, CVPR 2019

OccNet

Mescheder et al, CVPR 2019

DeepSDF

Park et al, CVPR 2019

slide-7
SLIDE 7

Compactness / low-poly / sharp details

Chen & Zhang, CVPR 2019

slide-8
SLIDE 8

Compactness / low-poly / sharp details

slide-9
SLIDE 9

Represent shapes as Binary Space Partitioning trees

*CSG operations – Constructive Solid Geometry operations

9

max-pooling min-pooling

slide-10
SLIDE 10

10

slide-11
SLIDE 11

11

slide-12
SLIDE 12

Network architecture

12

max-pooling min-pooling Dnxp Tpxc

D[: , 0]

Cnxc S*nx1

[: , 1] [: , 2] [: , 3] [: , 4] [: , 5] [: , 6] [: , 7] [: , 8] C[: , 0]

[: , 2] [: , 1]

  • 1

+1

shape surface 0

slide-13
SLIDE 13

Network architecture

13

max-pooling min-pooling

slide-14
SLIDE 14

Network architecture

14

max-pooling min-pooling

slide-15
SLIDE 15

Replacing max/min with weighted sum

15

slide-16
SLIDE 16

Replacing max/min with weighted sum

16

Dnxp Tpxc

D[: , 0]

Cnxc=relu(Dnxp)T S+nx1

[: , 1] [: , 2] [: , 3] [: , 4] [: , 5] [: , 6] [: , 7] [: , 8]

relu(Dnxp) relu(x) relu(1-Cnxc)

+ + + +

relu(1-Cnxc) 1cx1 relu(1-x)

1 ≈0

slide-17
SLIDE 17

Replacing max/min with weighted sum

17

Dnxp Tpxc

D[: , 0]

Cnxc=relu(Dnxp)T S+nx1

[: , 1] [: , 2] [: , 3] [: , 4] [: , 5] [: , 6] [: , 7] [: , 8]

relu(Dnxp) relu(x) [1-Cnxc][0,1] ≈relu(1-Cnxc)

+ + + +

[1-Cnxc][0,1] 1cx1 ≈relu(1-Cnxc) 1cx1

1 ≈0

slide-18
SLIDE 18

Replacing max/min with weighted sum

18

Dnxp Tpxc

D[: , 0]

Cnxc=relu(Dnxp)T S+nx1

[: , 1] [: , 2] [: , 3] [: , 4] [: , 5] [: , 6] [: , 7] [: , 8]

relu(Dnxp) relu(x) [1-Cnxc][0,1] ≈relu(1-Cnxc)

+ + + +

[1-Cnxc][0,1] Wcx1 ≈relu(1-Cnxc) 1cx1

1 ≈0

slide-19
SLIDE 19

Replacing max/min with weighted sum

19

Dnxp Tpxc

D[: , 0]

Cnxc=relu(Dnxp)T S+nx1

[: , 1] [: , 2] [: , 3] [: , 4] [: , 5] [: , 6] [: , 7] [: , 8]

relu(Dnxp) relu(x) [1-Cnxc][0,1] ≈relu(1-Cnxc)

+ + + +

[1-Cnxc][0,1] Wcx1 ≈relu(1-Cnxc) 1cx1

1 ≈0

[ [1 - relu(xPT) T ][0,1] W ][0,1]

slide-20
SLIDE 20

Replacing max/min with weighted sum

20

Dnxp Tpxc

D[: , 0]

Cnxc=relu(Dnxp)T S+nx1

[: , 1] [: , 2] [: , 3] [: , 4] [: , 5] [: , 6] [: , 7] [: , 8]

relu(Dnxp) relu(x) [1-Cnxc][0,1] ≈relu(1-Cnxc)

+ + + +

[1-Cnxc][0,1] Wcx1 ≈relu(1-Cnxc) 1cx1

1 ≈0

[ [1 - relu(xPT) T ][0,1] W ][0,1]

slide-21
SLIDE 21

Continuous weights Continuous weights Planes Convexes Output shape

Visualizing training - Initialization

21

slide-22
SLIDE 22

Visualizing training - Continuous phase

22

slide-23
SLIDE 23

Binary weights Min-pooling

Visualizing training – Discrete phase

23

slide-24
SLIDE 24

Binary weights Min-pooling

Visualizing training – Discrete phase

24

slide-25
SLIDE 25

Visualizing training – Discrete phase

25

slide-26
SLIDE 26

Toy dataset – 12 examples

26

slide-27
SLIDE 27

Toy dataset - Reconstruction

27

slide-28
SLIDE 28

Volumetric Primitives

Tulsiani et al, CVPR 2017

SuperQuadrics

Paschalidou et al, CVPR 2019

BAE-NET

Chen et al, ICCV 2019

BSP-NET (ours) Reconstruction ground truth

3D Reconstruction – Part Decomposition

28

slide-29
SLIDE 29

29

slide-30
SLIDE 30

Single View Reconstruction (RGB→3D)

ResNet BSP-Net ResNet BSP-Net ResNet BSP-Net ResNet BSP-Net ResNet BSP-Net

30

slide-31
SLIDE 31

AtlasNet (1 sphere)

Groueix et al, CVPR 2018

AtlasNet (25 patches)

Groueix et al, CVPR 2018

Input image

31

slide-32
SLIDE 32

Quantitative results

32

slide-33
SLIDE 33

Thank you. Q&A

33