SIGGRAPH 2022: Neural Geometry Processing

SIGGRAPH 2022勉強会で使ったスライドです。
https://siggraph.xyz/s2022/

Transcript

1. Neural Geometry Processing
   Tatsuya Yatagawa
   Aug. 27th, 2022
   SIGGRAPH 2022 Paper Reading Seminar
   

   View Slide

2. 6 TOG , 2 Conference Track
   l Dual Octree Graph Networks for Learning Adaptive Volumetric ...
   l DiffusionNet: Discretization Agnostic Learning on Surfaces
   l Neural Dual Contouring
   l Random Walks for Adversarial Meshes
   l PCEDNet: A Lightweight Neural Network for Fast and Interactive ...
   l Self-Sampling for Neural Point Cloud Consolidation
   l ImLoveNet: Misalignment Image-supported Registration Network ...
   l DSG-Net: Learning Disentangled Structure and Geometry ...
   Neural Geometry Processing (8 papers)
   

   View Slide

3. Dual Octree Graph Networks for Learning Adaptive
   Volumetric Shape Representations
   CNN ,
   

   View Slide

4. 1.
   2.
   ,
   3.
   MLP
   Dual Octree Graph
   

   View Slide

5. l , .
   l .
   

   View Slide

6. DiffusionNet: Discretization Agnostic Learning on
   Surfaces
   

   View Slide

7. :
   l 𝑡 ,
   DiffusionNet
   ,
   (= )
   l ,
   : , : +
   

   View Slide

8. ,
   l target
   l , ,
   .
   : , attention .
   , .
   

   View Slide

9. Neural Dual Contouring
   Dual Contouring
   

   View Slide

10. Marching Cubes ( : Dual Contouring, : Marching Cubes)
    Dual Contouring
    Dual Contouring
    ,
    Marching Cubes
    

    View Slide

11. l MC
    (MC )
    l MC ,
    ,
    : Neural Marching Cubes
    l ,
    (= )
    l
    

    View Slide

12. l NMC , DC
    l NMC , DC
    l NDC ,
    UNDC
    l DC
    l UNDC ,
    Neural Dual Contouring
    

    View Slide

13. l ( ) ,
    ,
    

    View Slide

14. Random Walks for Adversarial Meshes (Conf. Track)
    ,
    

    View Slide

15. ,
    (adversarial sample)
    : [Goodfellow et al., ICLR 2015]
    l
    

    View Slide

16. l RandomWalker [Lahav and Tal 2020]
    l
    : Imitating Network
    

    View Slide

17. l , random walk
    l Imitating Net ,
    :
    

    View Slide

18. l , (RandomWalker, MeshNet)
    l Imitating Network
    (Walk path Random Walker )
    

    View Slide

19. PCEDNet: A Lightweight Neural Network for Fast and
    Interactive Edge Detection in 3D Point Clouds
    

    View Slide

20. (Scale-Space Matrix)
    l Growing Least Squares [Mellado+ 2012]
    l , - , , ,
    .
    l 6 ,
    16 (=SSM) .
    PCEDNet (Point Cloud Edge Detection Network)
    l SSM (= 6 16 )
    .
    

    View Slide

21. l
    , ,
    ,
    l MLP, CNN
    l CNN ,
    ,
    l SSM 4-128
    ,
    l ,
    16
    4 16 128
    

    View Slide

22. Self-Sampling for Neural Point Cloud Consolidation
    ,
    Input Consolidated Input Consolidated
    

    View Slide

23. :
    l positive ( ) negative ( ) .
    l , (source)
    (target)
    , positive
    l
    

    View Slide

24. , .
    Input Consolidated Input Consolidated
    : , positive, negative
    , .
    

    View Slide

25. ImLoveNet: Misaligned Image-supported Registration Network
    for Low-overlap Point Cloud Pairs (Conf. Track)
    2
    

    View Slide

26. : , 2
    l (DGR [Choy+ 2020],
    PREDATOR [Huang+ 2021] etc.).
    l
    )
    

    View Slide

27. SOTA (PREDATOR, OMNet)
    : ,
    ,
    .
    

    View Slide

28. DSG-Net: Learning Disentangled Structure and
    Geometry for 3D Shape Generation
    (VAE)
    

    View Slide

29. :
    ,
    .
    :
    .
    

    View Slide

30. l 2
    +
    ( ) ( )
    l as-consistent-as-
    possible (9
    BBox ) [Guo+ 2019]
    l one-hot
    MLP
    l ACAP BBox
    ,
    .
    

    View Slide

31. (SDM-Net) , ,
    

    View Slide

32. 6 TOG , 2 Conference Track
    l Dual Octree Graph Networks for Learning Adaptive Volumetric ...
    l DiffusionNet: Discretization Agnostic Learning on Surfaces
    l Neural Dual Contouring
    l Random Walks for Adversarial Meshes
    l PCEDNet: A Lightweight Neural Network for Fast and Interactive ...
    l Self-Sampling for Neural Point Cloud Consolidation
    l ImLoveNet: Misalignment Image-supported Registration Network ...
    l DSG-Net: Learning Disentangled Structure and Geometry ...
    : Neural Geometry Processing
    : , .
    , (=
    ) .
    

    View Slide