I introduced Google's AIY Vision Kit at the Google Developers Group Boston Taking the Temperature of Tensorflow meetup. Google’s “AI Yourself” Vision Kit, now available at Target Stores, is a low-cost ($90), developer friendly development kit designed to make machine learning tangible for makers and STEM students. It includes pre-trained models for detecting mood, pets, meals, with the ability to load your own custom model.
5. JoyDetection
Demo
5/1/2018
cwh.consulting
The joy detection demo loads by default.
It does face detection and estimates your “Joy” score if you are smiling or frowning.
The default demo will illuminate the LED button according to the average score it sees
across all faces.
12. IntelMovidius
MA2450
Vision Processing Unit
5/1/2018
cwh.consulting
The most important piece of hardware is the intel Movidius VPU chip.
This is a specialized chip designed for running machine learning graphs at low power.
You can also buy this check as part of a USB dongle from Intel (Intel NCS) for use with Intel’s SDK.
The AIY Kit team built their own SDK and optimized it to work with Tensorflow.
13. Othergoodies
wewon’ttalk
abouttoday
Button with multi-color
LED
Privacy LED
Buzzer
Micro controller chip on
Bonnet with extra GPIO
pins, LEDs, and button
5/1/2018
cwh.consulting
The kit comes with lots of other interfaces and hardware options
15. Models
Face / Joy
detector
Bounding box
Joy score
Dog / Cat /
Human
Detector
Object class
Confidence
score
Bounding box
Dish Classifier
Food object
Confidence
score
Image Classifier
/ Google Image
Classifier
Image
classifier (1000
classes?)
Confidence
score
Nature
Explorer
Image
classification of
4,080 different
species
5/1/2018
cwh.consultingGoogle maintains several computer vision (CV) models that can be loaded on to the kit.
Several of these were recently added and hopefully more will come soon.
16. Process
1. Develop the
model
2. Train 3. Freeze graph &
compile
4. Deploy
5/1/2018
cwh.consultingModels need to be trained outside of the kit.
Then they can be compiled for the AIY Vision Kit and deployed there.
17. Custommodels
Tensorflow Model Compiler
Only works on x86 64 Ubuntu
Lots of limitations on the model – specific operators and
configuration require
Build your own
Retrain an existing model
5/1/2018
cwh.consulting
./bonnet_model_compiler.par
--
frozen_graph_path=cat_detector_interference_graph_aiy/frozen_inference_graph.p
b
--output_graph_path=cat_detector.binaryproto
--input_tensor_name="input"
--output_tensor_names="concat,concat_1"
--input_tensor_size=160
--debug
There are caveats compiling to the AIY Vision Kit.
I have not been able to get this to work (or seen someone else do this), but I think we are very close – see
https://github.com/google/aiyprojects-raspbian/issues/314
18. Demos
AIY Vision Kit
5/1/2018
cwh.consulting
Cat detection using the built in cat/dog/human detector: https://youtu.be/NguWj9K1MJs
Face detection: https://webrtchacks.com/wp-content/uploads/2018/02/face_detection_camera.py-preview.mp4?_=1
19. What does it all mean?
2 things…
5/1/2018
cwh.consulting
22. EdgeML
Examples of Computer Vision on
the Movidius
5/1/2018
cwh.consulting
Here are some real products that use the Movidius chip for running machine learning on-device
24. STEM
5/1/2018
cwh.consulting
2. The kit is designed to expose high-schoolers to using AI technologies, not for PhD’s
Google is trying to make ML tangible for a large audience.