6. Google I/O 2018 Session
“ML models + IoT data
= a smarter world”
https://www.youtube.com/watch?v=avxpkFUXIfA
7. Regulations of this technical demo
● Adopt Cloud IoT Core
● Utilize ML model & Cloud ML Engine (training and inference)
● ML model inference on edge device is out of scope
(It’s included in Laurence’s part of the session)
9. Demo dashboard
Camera Image +
Object Detection Results
Recommended recipe
according to cart contents+α
Information about the foodstuffs
required for the recipe
14. 3/27
initial offer from
kaz-san
〜4/5
Draw rough blueprint
of demo
〜4/13 4/27
Pivot!!
5/9
Google I/O Session
Device (Raspberry Pi)
software and ML model
development deadline.
15. 3/27
initial offer from
kaz-san
〜4/5
Draw rough blueprint
of demo
〜4/13
Pivot!!
4/27
Device (Raspberry Pi)
software and ML model
development deadline.
5/9
Google I/O Session
GW
(Japanese
Holidays)
kaz-san’s
departures to US
16. In fact only 2 weeks are available for
development
● Choose and purchase hardware setup (Raspberry Pi, camera module,
touch display etc...)
● Purchase imitation foodstuff (onion, tomato, potato, etc...)
● Device setup & demo application development
● Capture images for ML model training (Object Detection)
● Thinking of recipes
● Collect images for each recipe
● Generate dummy foodstuff sequence of shopper’s pickup.
● Training ML models (Object Detection & Next item prediction)
● Dashboard application
22. ML models
● Based on Tensorflow Object Detection API
https://github.com/tensorflow/models/tree/master/research/object_detection
○ SSD + mobilenet v1
○ Transfer learning (upon COCO-trained model)
● Next shopping item prediction
○ 1D-Conv + MLP model
ML models are trained using
28. ML inference should be executed on cloud
● We should think of reasons justify cloud inference
○ Edge inference was optimum in most scinario
29. The demo should be attractive and easy to
understand at glance
● “Small sensor device” is typical and realistic IoT system scenario
○ Temperature, Air pressure, Air dust etc..
● Attractive visualization is hard to develop
● Image recognition is intuitive and easy to visualize!
(But is it IoT thing?)
30. All images should be licensed under CC0
https://github.com/groovenauts/SmartShoppingNavigator/issues/11
● Collecting CC0 or Public Domain attractive images is time consuming stuff
● Deeply thanks to pixabay (https://pixabay.com/)
31. Recipes should be those which
US people are familiar with
● Culture gap?
● There are less typical cuisine with common name (really?)
● Cookpad (https://cookpad.com/us) show variety of recipes
But they are like “Bacon and potato”, “○○’s speciality”...