When I was in college (around 2010), I worked with OpenCV and facial detection quite a bit, I always found it very interesting. I wanted to build a recognition platform on top of this, but I quickly determined I was in way over my head…
Neural networks have come a long way since then.
Since moving into our home, I have been trying to find ways of gathering data. Temperatures inside and outside, moisture levels in the garden, todo list data, camera systems with motion detection, etc. I have been doing this in the hope that one day, I can create a neural networks to help with things around the house.
Recently, I have had a breakthrough in this area.
After noticing that I was accidentally storing hundreds of gigabytes of motion detection videos, I realized the treasure this could be!
Finding the interesting bits
Since I peruse the tech blogs every day, I remembered seeing some new work in the image tagging world. After some searching, I found it again, YOLO. After building this library, downloading a previously trained model, and learning a bit of python, I was labeling objects in my videos!
At first, I naively thought there was no way this could work for my use case. Each tag took around 30 seconds to calculate and I was potentially going to have thousands of frames thrown at this thing.
NO. I went a much simpler route that has bene working well.
Only a few frames
I decided there was no reason to analyze entire videos worth of frames. My motion detection was already doing the hard work for me. I decided to break out ffmpeg and extract a few frames of the detection video.
This greatly simplified my problem. Now all I had to do was tag a few images for each detection video.
I quickly wrote a simple python server that accepted an POST request with an image, and used YOLO to spit out a list of labels. Each which had bounding box data… this was getting exciting.
Tagging
Just for fun, I whipped up a quick react app which allowed me to swipe through all my tagged objects and give them real labels.
Personal Recognition
To be honest, the labels were nice, but those were not what I was after. I wanted to know when my wife walks through the gate, if my dog is barking at the fence, or if a friend comes over. That was my goal.
I went back to a library I had tried before and failed with. ConvNetJS I had remembered he made a demo which could learn to recognize objects from a dataset with 90% accuracy after just a few minutes.
Here is the demo.
I now had a large amount of raw training data, and a potential library to use.
I knew my raw data needed to be cleaned up and normalized for it to run through a network. I kept it simple, I cropped out the detected boundary of the image then used a contain algorithm to resize them to 64x64. In the future I want to also save a horizontal flipped version of the image so I can double my training data.
It took a fews day to figure out ConvNetJS. At first I was only running the data through once per image and expecting it to work 🤦.
I went back to the demo and realized he was throwing random input images at the network and teaching it the correct label, while minimizing the "classification loss". I coded up something very similar, but without the fancy graphs. It was just a terminal app that took all my normalized images and personal labels and trained them randomly over a set amount of iterations.
After tweaking the network training configuration a few times, my training normally takes around 30 mins, and performs quite well.
I now get emails throughout the day of recognized people, dogs, cars, etc.
Still more work to do, but I am very happy with where it is right now.