Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Building an intelligent Pullup-Classifier - Iterations, Learnings, Thoughts

57 views

Published on

OMM TECHtalks #1
www.tech-talks.eu

Published in: Software
  • Be the first to comment

  • Be the first to like this

Building an intelligent Pullup-Classifier - Iterations, Learnings, Thoughts

  1. 1. OMM Solutions TECHtalk #1 www.omm-solutions.de 30.03.2017 < OMM Solutions GmbH > 1 www.tech-talks.eu
  2. 2. Einmal im Monat ist TECHtalk Zeit! First come first served! < OMM Solutions GmbH > 2
  3. 3. Talk: Building an intelligent Pullup-Classifier Speaker: Olaf Horstmann 3< OMM Solutions GmbH >
  4. 4. 30.03.2017 < OMM Solutions GmbH > 4
  5. 5. • should COUNT the pullups • should detect WHO • should classify the QUALITY The Goal 30.03.2017 < OMM Solutions GmbH > 5
  6. 6. Image from: http://www.instructables.com/id/Pull-Up-Counter/ Existing solutions 30.03.2017 < OMM Solutions GmbH > 6
  7. 7. 30.03.2017 < OMM Solutions GmbH > 7
  8. 8. ✓ COUNT × WHO × QUALITY Can we use it? 30.03.2017 < OMM Solutions GmbH > 8
  9. 9. The Pressure-Sensor Iteration #1 30.03.2017 < OMM Solutions GmbH > 9
  10. 10. Let‘s try it! 30.03.2017 < OMM Solutions GmbH > 10
  11. 11. OKAY… 30.03.2017 < OMM Solutions GmbH > 11
  12. 12. The Ultrasonic-Sensor Iteration #2 30.03.2017 < OMM Solutions GmbH > 12
  13. 13. 30.03.2017 < OMM Solutions GmbH > 13
  14. 14. 30.03.2017 < OMM Solutions GmbH > 14
  15. 15. • Only works on perpendicular angles (+/-15°) • Only works on „flat“ surfaces (hair or clothes = bad) • You can hear it when it is quiet BUT… 30.03.2017 < OMM Solutions GmbH > 15
  16. 16. An optical sensor Iteration #3 30.03.2017 < OMM Solutions GmbH > 16
  17. 17. 30.03.2017 < OMM Solutions GmbH > 17
  18. 18. 30.03.2017 < OMM Solutions GmbH > 18
  19. 19. Success 30.03.2017 < OMM Solutions GmbH > 19
  20. 20. • Cost more than the other sensors • Inaccurate/slow > 2m • Might not work very well in direct sunlight (did not test) Any Drawbacks? 30.03.2017 < OMM Solutions GmbH > 20
  21. 21. • Collected data from different people • Prepate the data • Implement an AI to classify the data Now what? 30.03.2017 < OMM Solutions GmbH > 21
  22. 22. Collecting/Comparing Data 30.03.2017 < OMM Solutions GmbH > 22
  23. 23. • Algorithm to detect individual pullups (=> count) • Label each data-set manually (=> who & quality) Data-Preparation 30.03.2017 < OMM Solutions GmbH > 23
  24. 24. Data-Preparation 30.03.2017 < OMM Solutions GmbH > 24
  25. 25. • Throw in the raw data (~200 pullups) • Train it • See what happens • => accuracy ~60% With Tensorflow AI Iteration #1 30.03.2017 < OMM Solutions GmbH > 25
  26. 26. • More data (a lot more!) • Adjust the network (not really better) • Use meta-data instead of raw data How can we improve this? 30.03.2017 < OMM Solutions GmbH > 26
  27. 27. • Using ~15 different meta-data-points • => accuracy ~85% AI Iteration #2 30.03.2017 < OMM Solutions GmbH > 27
  28. 28. • Normalize data (high values are bad!) • …and more data! …but there‘s still room for improvements It‘s Okay… 30.03.2017 < OMM Solutions GmbH > 28
  29. 29. • Normalized data • => accuracy at +90% AI Iteration #3 30.03.2017 < OMM Solutions GmbH > 29
  30. 30. „The Product“ 30.03.2017 < OMM Solutions GmbH > 30
  31. 31. • Data preparation is probably the most important step • Visualizing the data helps analyzing it • Less data can be compensated by transforming the data first • Having more classes can help with the accuracy of the whole model • More data is always better Learnings 30.03.2017 < OMM Solutions GmbH > 31
  32. 32. Vielen Dank für Eure Aufmerksamkeit! 32< OMM Solutions GmbH >
  33. 33. Ihr persönlicher Ansprechpartner Fragen oder Interesse? < OMM Solutions GmbH > 33 Olaf Horstmann Technology & Innovation OMM Solutions GmbH Vor dem Lauch 4 70567 Stuttgart Germany oh@omm-solutions.de +49 (0)711 75 86 46 04
  34. 34. 34< OMM Solutions GmbH > www.omm-solutions.de OMM Solutions GmbH Vor dem Lauch 4 70567 Stuttgart Geschäftsführer Martin Allmendinger Malte Horstmann Olaf Horstmann Kontakt Telefon: +49 711 6747 051-0 E-Mail: info@omm-solutions.de Umsatzsteuer-ID: DE295716572 Sitz der Gesellschaft: Stuttgart Amtsgericht Stuttgart, HRB 749562 Impressum

×