$30 off During Our Annual Pro Sale. View Details »

【授業スライド】AnonyMask

 【授業スライド】AnonyMask

慶應義塾⼤学 B4 和田唯我 / Yuiga Wada
授業『機械知能』のスライドです. (院先取り)

- About me
- Blog

Yuiga Wada (和田唯我)

October 05, 2022
Tweet

More Decks by Yuiga Wada (和田唯我)

Other Decks in Technology

Transcript

  1. AnonyMask
    XXXXXX Yuiga Wada YYYYY Seitaro Otsuki ZZZZZZ Yui Iioka
    Machine Intelligence Demonstration (Team-7)
    1
    https://github.com/OtsuKotsu/Anonymask

    View Slide

  2. AnonyMask : Remove Corporate Logo naturally
    ● Background
    ○ Huge cost to hide the logos
    ○ Existing approaches result in unnatural images!
    ● Our product : AnonyMask
    ○ Detects the logo by the object detection model
    ○ Masks it out and repaints masked region naturally
    2

    View Slide

  3. Method : AnonyMask
    ● Overall flow
    1. Object detection
    2. Masking & Reconstruction
    3. Smoothing & Improving Resolution
    3

    View Slide

  4. Method : AnonyMask
    YOLOv5
    Masked Auto Encoder
    SwinIR
    4

    View Slide

  5. Results : AnonyMask generates remarkably natural images 👀
    5

    View Slide

  6. Demonstration
    https://drive.google.com/file/d/1n0Ir2d-zZ9MazjbOpYz54ibGo540elnH/view?usp=sharing
    6

    View Slide

  7. Quantitative Result
    Recall Precision F-Score [email protected] [email protected]:0.95
    val. 0.656 0.789 0.716 0.732 0.507
    test 0.653 0.779 0.710 0.728 0.500
    7

    View Slide

  8. Thank you for listening
    Thank you for listening !
    https://github.com/OtsuKotsu/Anonymask
    8

    View Slide

  9. Appendix : Method (1/4)
    ● YOLO v5
    ○ Detects the position and type of objects
    ○ Metrics : mAP (mean Average Precision)
    ● Overall flow
    1. Object detection
    2. Masking & Reconstruction
    3. Smoothing & Improving Resolution
    9

    View Slide

  10. Appendix : Method (2/4)
    ● Overall flow
    1. Object detection
    2. Masking & Reconstruction
    3. Smoothing & Improving Resolution
    ● MAE [He+, CVPR22]
    ○ Masks the logo and Reconstructs from
    surrounding pixels
    ○ Metrics : MSE
    10

    View Slide

  11. Appendix : Method (3/4)
    ● Overall flow
    1. Object detection
    2. Masking & Reconstruction
    3. Smoothing & Improving Resolution
    ● Gaussian filter
    ○ Erases noise in an image
    ○ Blurs images
    11

    View Slide

  12. 12
    Appendix : Method (4/4)
    ● Overall flow
    1. Object detection
    2. Masking & Reconstruction
    3. Smoothing & Improving Resolution
    ● SwinIR [Liang+, ICCV21]
    ○ Restores high-quality images from
    low-quality images

    View Slide

  13. Appendix : Dataset
    13
    QMUL-OpenLogo
    ● 27,083 images from 352 logo classes
    ● Object-level Annotations
    Our split :
    https://hangsu0730.github.io/qmul-openlogo/
    #Examples
    Training set 18752
    Validation set 4165
    Test set 4166

    View Slide

  14. Appendix : demo details (1/4)
    YOLOv5
    MAE
    SwinIR
    Server
    Client
    https://XXXX:port
    14

    View Slide

  15. Appendix : demo details (2/4)
    15

    View Slide

  16. Appendix : demo details (3/4)
    16

    View Slide

  17. Appendix : demo details (4/4)
    17

    View Slide

  18. Reference
    1. K. He, X. Chen, S. Xie, et al., “Masked autoencoders are scalable vision learners,” CVPR,
    pp.16000–16009, 2022.
    2. J. Liang, J. Cao, G. Sun, et al., “Swinir: Image restoration using swin transformer,” CVPR,
    pp.1833–1844, 2021.
    3. G. Jocher, A. Chaurasia, A. Stoken, et al., “ultralytics/yolov5: v6.1 - TensorRT, TensorFlow Edge TPU
    and OpenVINO Export and Inference,” 2022. https://doi.org/10.5281/zenodo.6222936
    18

    View Slide