GISA2018_Mapillary

* 地理情報システム学会第27回学術研究発表大会
* Mapillary＆OpenStreetMapによるマイクロマッピング

Transcript

1. /2 8
   
   
   !
   
   10 M
   

   View Slide

2. /0 12 8
   -- 0 : 1
   A B 4 C 1 :2
   

   View Slide

3. 01 / / 2 4 8
   
   
   
   
   

   View Slide

4. 01 / / 2 5 8
   https://www.mapillary.com/app/
   

   View Slide

5. 01 / / 2 6 8
   
   

   View Slide

6. 01 / / 2 8 7
   
   
   

   View Slide

7. /0 12 8
   
   
   

   View Slide

8. 01 / / 2 89
   Source: https://taginfo.openstreetmap.org/keys/mapillary#overview
   

   View Slide

9. /0 12 8
   -
   

   View Slide

10. /0 12 8
    # # #
    https://mapillary.github.io/mapillary_greenhouse/
    

    View Slide

11. /0 12 8
    https://mapillary.github.io/mapillary_greenhouse/global-challenge/q2-2018/
    Score = (total unique KM contributed + (images/1000)) * (1 + (count of local participants/10))
    

    View Slide

12. 01 / / 23 8
    4
    

    View Slide

13. 01 / / 2 4 8
    Mapillary images are available under the Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)
    

    View Slide

14. 01 / / 2 5 8
    
    
    https://www.mapillary.com/pricing
    

    View Slide

15. 01 / / 2 6 8
    
    
    http://openaccess.thecvf.com/content_ICCV_2017/papers/Neuhold_The_Mapillary_Vi
    stas_ICCV_2017_paper.pdf
    

    View Slide

16. 01 / / 2 8 7
    
    
    The Mapillary Vistas Dataset for Semantic Understanding of Street Scenes
    doi: 10.1109/ICCV.2017.534
    

    View Slide

17. /0 12 8
    - - - - ( -
    ( - - )
    https://research.mapillary.com/lsun.html
    

    View Slide

18. 01 / / 2 89
    Mapillary Vistas Dataset
    25,000 & 100
    
    https://research.mapillary.com/
    

    View Slide

19. /0 12 8
    0 2 1 ,
    •Road markings: increased from 6 classes to 21 classes including speed limits and direction
    arrows, which are essential classes for building navigation and HD maps.
    •Driveway: a new class that we’ve introduced to help enhance data extraction for last-mile
    •Barriers: increased from 6 to 11 classes. Detailed refinement of the barrier classes is the first
    step to enhance the perception capacity for autonomous driving under different scenarios.
    •Signages: the billboard class is now refined to advertisement, storefront, and general
    information signages. This facilitates automatic and in-depth location-aware data extraction
    for marketing, business, and point-of-interest analysis.
    •Traffic light states: dynamic properties are associated with static objects. In this case, we
    have refined 100K traffic light annotations with different state
    

    View Slide

20. /0 12 8
    http://www.robustvision.net/
    

    View Slide

21. /0 12 8
    https://blog.mapillary.com/update/2018/06/14/robust-cv.html
    

    View Slide

22. 01 / / 23 8
    :
    https://arxiv.org/abs/1712.02616
    

    View Slide

23. 01 / / 2 4 8
    In-Place Activated BatchNorm for Memory-Optimized Training of DNNs
    https://arxiv.org/abs/1712.02616
    

    View Slide

24. 01 / / 2 5 8
    3/ ) ( 0 . 3 .
    https://github.com/mapillary/mapillary-js
    

    View Slide

25. 01 / / 2 6 8
    https://blog.mapillary.com/update/2018/06/12/mapillary-for-organizations.html
    

    View Slide

26. . 788 / 6 1 66 . MQ e V e W
    3 . (
    https://github.com/enricofer/go2mapillary
    • 0 1 ( lnx
    i WM
    G
    • tpJO V
    i W
    • 0 tp 0 1 M
    IG 0365276
    • . 788 1
    b Jr
    • aS Vw
    Thanks! Christopher Beddow
    

    View Slide

27. 0 0 C . 0 RSe Wle cm
    . . / .
    • Fdmh m O
    r . 0
    F Ok fF F
    • 0 0 C . 0:
    / 8 : 2 0 ::
    P 18
    • yv Stw O
    nsoMp
    • mai fF gm
    4 0:0 1
    0 1
    Thanks! Christopher Beddow
    

    View Slide

28. /8244/9 81 . 911 /8 M PS a
    • a I
    . C c
    • 1IS
    198/ Ie
    • 20 /9
    • IOa i
    nI j
    • in p r
    om l AM
    . . / .
    Thanks! Christopher Beddow
    

    View Slide

29. 01 / / 23 8
    
    
    

    View Slide

30. 01 / / 23 8
    Thank you!
    tosseto@csis.u-tokyo.ac.jp
    http://researchmap.jp/tosseto
    https://speakerdeck.com/tosseto
    

    View Slide