Image classification in real world : Car rental damage report

Transcript

1. Image classification in real world Car rental damage report
5. TOP 7 Amazing Tips About Machine Learning for Dummies (aka iPhone developers)
7. Image classification Folders train Unkown image Classes
8. #1 ML is not only for data-scientists
11. My Dataset 41 269 pictures 62,63 Go Same cars Same set of colors Same luminosity (no light) Already classified by users
14. #2 You already have a dataset or you should
15. UX proposition Is it a Front alloy wheel ? Yes sure ! No it’s not
16. Create ML Top 1 - From 1 34 %
17. Wheel Tire Fender Gas trap
18. Create ML Top 1 - From 1 34 % Top 1 - From 5 91 %
19. UX proposition v2 Something else
20. Wheel Roof Fender Gas trap
21. Create ML Top 1 - From 1 34 % Top 1 - From 5 91 % Top n - From n 78 %
22. #3 Don’t trust results Use it as a hint
26. Base Merge Classes Top 1 - From 1 34 % 37 % Top 1 - From 5 91 % 93 % Top n - From n 78 % 81 %
27. #4 Merge classes If you can’t see it, ML can’t see it
28. Partner app
29. Base Partner data Top 1 - From 1 37 % 34 % Top 1 - From 5 93 % 89 % Top n - From n 81 % 75 %
30. Base Partner data Partner corrected data Top 1 - From 1 37 % 34 % 41 % Top 1 - From 5 93 % 89 % 94 % Top n - From n 81 % 75 % 84 %
31. #5 Clean your data Size of dataset don’t always matter
33. Inception v3 - MobileNet Inception v3 MobileNet Create ML Top 1 - From 5 91 % 90 % 91 % Top n - From n 79 % 18 % 78 % Size 87 Mo 2.2 Mo 1.5 Mo Training time 2h41 2h38 1h50
34. #6 Use Create ML It’s really cool
35. #1 ML is for mobile developer too #2 You already have a dataset #3 Don’t trust the result #4 If you can’t see it, ML can’t see it #5 Clean your data #6 Use Create ML #7 Drive safe
36. #1 ML is for mobile developer too #2 You already have a dataset #3 Don’t trust the result #4 If you can’t see it, ML can’t see it #5 Clean your data #6 Use Create ML #7 Drive safe
37. Mathieu Hausherr @mhausherr