Simple Learning is the result of my efforts to create a custom solution to identify raccoons at our cat bowl. I quickly realized that raccoons match a set of categories which are not called out in the 1000 list of categories trained by the SqueezeNet model. I saw that I could generalize the training by letting the camera run on a background and collect features. A detection run then identifies new objects which are not in the background set. If the probability is high enough, it sends an SNS message by SMS text message or email along with the URL to the user.
I also found a way to 'unleash' deep lens by using a portable AC Power Supply and using my phone as a hotspot.
Finally, I use Alexa to control Deep Lens Simple Learning. I have an intent to start training and one to start detection. An S3 bucket acts as a state manager between Deep Lens and Alexa. This allows voice control of the device. The user can also retrieve the Current State of Simple Learning. The bucket also acts as an image repository for transmission to the user.
What it does
It uses Alexa to start background training and detection. The camera can be placed anywhere. If an outlet is not close by, then the AC Power Supply meets the need. It detects objects which were not in the background training set and alerts the user. Thresholds can be adjusted for refinements.
How I built it
I built it using a set of Lambda functions, Alexa, and many other AWS services.
Challenges I ran into
Time. After Dog Park, this one seemed logical. It generalizes the whole training aspect of Deep Lens.
Accomplishments that I'm proud of
What I learned
That I can actually build something that seemed too difficult when I started. I learned a lot.
What's next for Deep Lens Simple Training
There is a lot more to come!