1 of 51

3D-R N

Christopher B. Choy, Danfei Xu*, JunYoung Gwak*, Kevin Chen, Silvio Savarese

A unified approach for single and multi-view 3D object reconstruction

2

2

Computational Vision

& Geometry Lab

2 of 51

3D Reconstruction

  • Navigation
  • Robot interaction, manipulation
  • 3D object prototyping
  • 3D printing

Computational Vision

& Geometry Lab

3 of 51

3D Reconstruction

  • Depth based method [Eigen et al., Saxena et al., etc]
  • Model based methods [Kar et al., Aubry et al., Choy et al., etc]
  • Structure from Motion (SfM) [Haming et al., Fuentes-Pacheco et al.]
  • Multi view Stereo [Seitz et al., Anwar et al., etc]

Computational Vision

& Geometry Lab

4 of 51

Reconstruction

  • Lambertian and non-uniform albedo
    • Non-reflective
    • Rich of non-homogeneous textures
  • Dense viewpoints (small baseline)

Computational Vision

& Geometry Lab

5 of 51

Reconstruction

without assumptions

Computational Vision

& Geometry Lab

6 of 51

Reconstruction

  • Shape Prior / Data-driven method
  • Single View Reconstruction (Kar et al., Aubry et al., etc)

  • Multi View Reconstruction (Bao et al.)
    • General lighting condition

without assumptions

    • Localize parts
    • Boundaries

    • Dense viewpoints

Computational Vision

& Geometry Lab

7 of 51

3D-Recurrent Reconstruction Neural Network

  • Data Driven Reconstruction
    • [Saxena et al., Hoiem et al., Vincente et al., Kar et al.]
  • Recurrent Neural Network
    • Sequence of images
    • Probabilistic voxel occupancy map
  • 3D-Convolutional LSTM
    • Locality
    • Regularize connection
    • Selective Update

Computational Vision

& Geometry Lab

8 of 51

Recurrent Neural Network

[Christopher Olah] Understanding LSTM Networks, http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Computational Vision

& Geometry Lab

9 of 51

Long Short Term Memory

[Christopher Olah] Understanding LSTM Networks, http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Computational Vision

& Geometry Lab

10 of 51

Computational Vision

& Geometry Lab

11 of 51

3D Convolutional LSTM

Computational Vision

& Geometry Lab

12 of 51

Computational Vision

& Geometry Lab

13 of 51

Computational Vision

& Geometry Lab

14 of 51

Computational Vision

& Geometry Lab

15 of 51

Computational Vision

& Geometry Lab

16 of 51

3D Convolutional LSTM

Left front top

Left front top

Computational Vision

& Geometry Lab

17 of 51

3D Convolutional LSTM

  • Locality
  • Regularize connection
  • Selective Update

Computational Vision

& Geometry Lab

18 of 51

3D Convolutional LSTM

  • Locality
  • Regularize connection
  • Selective Update

Computational Vision

& Geometry Lab

19 of 51

3D Convolutional LSTM

  • Front/Side
  • Back

Computational Vision

& Geometry Lab

20 of 51

Computational Vision

& Geometry Lab

21 of 51

Computational Vision

& Geometry Lab

22 of 51

Training

  • ShapeNet
    • 50k CAD models
    • Render from arbitrary views
    • Random number of images w/ random order
    • Random background, translation
  • Voxel-wise cross entropy loss

Computational Vision

& Geometry Lab

23 of 51

Experiment I: Multi View Stereo vs Ours

Experiment II: Ebay Image Reconstruction

Experiment III: PASCAL 3D Single-View Reconstruction

Experiment IV: ShapeNet Multi-View Reconstruction

Computational Vision

& Geometry Lab

24 of 51

Experiments

Experiment I: Multi View Stereo vs Ours

Computational Vision

& Geometry Lab

25 of 51

  • SfM Limitations
    • Texture level
    • Number of views

Computational Vision

& Geometry Lab

26 of 51

Computational Vision

& Geometry Lab

27 of 51

Computational Vision

& Geometry Lab

28 of 51

  • Resolution
    • Input
    • Output

MVS

Ours

20

30

40

Computational Vision

& Geometry Lab

29 of 51

Experiment II: Ebay Image Reconstruction

Computational Vision

& Geometry Lab

30 of 51

Computational Vision

& Geometry Lab

31 of 51

Computational Vision

& Geometry Lab

32 of 51

Computational Vision

& Geometry Lab

33 of 51

Experiment III: PASCAL 3D Single-View Reconstruction

Computational Vision

& Geometry Lab

34 of 51

Training

  • PASCAL 3D+
    • Augmented PASCAL images with 3D CAD models
    • 3D Intersection over Union

Computational Vision

& Geometry Lab

35 of 51

Computational Vision

& Geometry Lab

36 of 51

Computational Vision

& Geometry Lab

37 of 51

Computational Vision

& Geometry Lab

38 of 51

Computational Vision

& Geometry Lab

39 of 51

Experiment IV: ShapeNet Multi-View Reconstruction

Computational Vision

& Geometry Lab

40 of 51

Computational Vision

& Geometry Lab

41 of 51

Computational Vision

& Geometry Lab

42 of 51

Computational Vision

& Geometry Lab

43 of 51

Thank you

Computational Vision

& Geometry Lab

44 of 51

Computational Vision

& Geometry Lab

45 of 51

Computational Vision

& Geometry Lab

46 of 51

3D Convolutional LSTM

Computational Vision

& Geometry Lab

47 of 51

Results

  • Multi-view stereo tends to fail when
    • Viewpoints are sparsely positioned
    • Objects are textureless
    • Not enough views
  • Fig. 8 from paper

Computational Vision

& Geometry Lab

48 of 51

Computational Vision

& Geometry Lab

49 of 51

Computational Vision

& Geometry Lab

50 of 51

Computational Vision

& Geometry Lab

51 of 51

3D-Convolutional LSTM

  • 4D tensor
  • No output gate
  • 3D Convolution

Computational Vision

& Geometry Lab