learning to shade hand- drawn sketches39.96.165.147/seminar/yuhan_200405.pdf · learning to shade...

32
Learning to Shade Hand- drawn Sketches CVPR 2020 Oral Presenter: Yu Han 2020/04/05 1

Upload: others

Post on 18-Oct-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

  • Learning to Shade Hand-drawn Sketches

    CVPR 2020 Oral Presenter: Yu Han

    2020/04/05

    1

  • Outline • Authors • Background • Network • Experiment • Conclusion

    2

  • Qingyuan Zheng* · University of Maryland

    3

  • Zhuoru Li* · Project HAT

    4

  • Adam Bargteil · Assistant professor at University of Maryland

    · Phd at UCB · Postdoc at CMU Graphics Lab

    Research Interest: · Computer graphics and animation, especially using physics-based animation

    5

  • Outline • Authors • Background • Network • Experiment • Conclusion

    6

  • Background• Pix2pix

    7

    Image-to-Image Translation with Conditional Adversarial Networks. In CVPR, 2017.

  • Background• DeepNormal

    8

    Deep Normal Estimation for Automatic Shading of Hand-Drawn Characters. In ECCV, 2018.

  • Background• Sketch2Normal

    9

    Interactive Sketch-Based Normal Map Generation with Deep Neural Networks. In SIGGRAPH, 2018.

  • Background• Squeeze and Excitation(SE) Block

    10

    SE Net:Squeeze-and-Excitation Networks. In CVPR, 2018.

  • Outline • Authors • Background • Network • Experiment • Conclusion

    11

  • Network

    12

  • Data

    13

    · 1160 paired hand-drawn line drawings and shadows · Tagged with a lighting direction · 8x3+2=26 directions

  • Network Architecture 1. Generative Network 2. Discriminator Network 3. Loss Function

    14

  • Generative Network

    15

    Two parts: • Shape net: encodes the underlying 3D structure from 2D sketches • Render net: renders artistic shadows based on the encoded structure

  • Generative Network

    16

    Render net: • Self-attention: enhance the visual reasoning 3D structure from 2D sketches. • SE blocks: filter out unnecessary features • Extract two supervision side outputs, s1 and s2

  • Network Architecture 1. Generative Network 2. Discriminator Network 3. Loss Function

    17

  • Discriminator Network

    18

    • Self-attention: make discriminator sensitive to the distant features

  • Network Architecture 1. Generative Network 2. Discriminator Network 3. Loss Function

    19

  • Loss Function

    20

  • Outline • Authors • Background • Network • Experiment • Conclusion

    21

  • Experiment

    22

    • Shift, zoom in/out, and rotate to augment the dataset • Input: 320x320 • 80 000 iterations

  • Experiment

    23

  • Experiment

    24

  • Experiment

    25

  • Experiment

    26

  • User Study

    27

    All people.

    People with drawing experience.

  • Ablation Study

    28

  • Ablation Study

    29

    Calculate the first two orders of the distance between distributions in Gaussian space.

  • Outline • Authors • Background • Network • Experiment • Conclusion

    30

  • Conclusion • New dataset

    • Propose a network that “understands” the structure and 3D spatial

    relationships implied by line drawings and produces highly-detailed

    and accurate shadows

    31

  • Thank you

    32