We do Deep-Learning.
Oct 14, 2016 • Richard Johnson
If you’ve had the fortune to bump into to any of us here are Silverpond, you’ll know that in addition to being generally geeky, we’re kind of into deep learning. It should be no surprise then that when a couple of us found ourselves with some free time on our hands, we decided to combine some of the various bits of hardware we had lying around the office with our love of deep learning to create something cool.
And this is what we came up with, a little system that will take photos of the office and count the number of people it can see:
The setup is reasonably simple, a Raspberry Pi takes a photo of the office through a digital SLR and uploads it to a bucket on AWS S3. Once a day, we wake an EC2 instance which processes all of the uploaded photos and does the heavy lifting of locating the people in the photo.
The first step was to get the Pi talking to the digital SLR. While there are a number of open-source tools available which will allow you to do this, we instead used some software we had previously developed to get around some of the performance and stability issues we had encountered with the open-source tools. From there we simply used the AWS CLI tools to upload the captured image into S3.
We didn’t particularly want to have an EC2 instance running constantly doing very little but processing the odd image every now and then, so we had a quick look for other options. At first glance it seemed that AWS Lambda would be perfect for this little task; we have the Lambda triggered whenever a file was uploaded into S3 and not have to worry about servers at all. Unfortunately and understandably there are some reasonably tight resource restrictions on AWS Lambda, which made this infeasible. The big one for us was the limit of 50MB in a deployment package; deep learning requires a hefty set of supporting libraries, and TensorFlow by itself easily pushes your package into the hundreds of MB.
So, as tempting as it was to jump down this rabbit hole and start looking at the various ways we could trim things back (we have previously had success running deep learning models on low-spec devices), we instead decided to go with a more traditional approach and use a normal EC2 instance which we would simply spin up when required. This is achieved through a small Lambda function which is triggered daily through a CloudWatch scheduled event to start the stopped instance. The server itself has a cron job configured to run on reboot that will download all the images, process them, upload the results and then turn itself off.
This is a reasonably simple pipeline but it would be straightforward to turn this into a more scalable system that could handle a considerable number of images. One approach would be to place the images onto a distributed queue service such as SQS; if we detect that the queue is not draining quickly enough, worker servers could be automatically launched to assist with the processing, scaling back down again once the load drops off.
Now that we had the basic supporting infrastructure in place, it was time to have a look at the deep learning side of things. Recently image processing has become a massively improved area of machine learning thanks to the publication of, and ongoing improvements to Google’s “inception” convolutional neural network. Through some data wizardry we can take this existing work and bend it to perform different tasks; in this case we wanted it to locate pedestrians in the photo. A little additional decoration was added via a bounding box outlining the areas the model thinks belongs to a “human”.
Once we had this magic up and running, it was simply a matter of plugging it into our simple AWS pipeline and viola, a pedestrian detector.
While we put together this little demo in a couple of days, we think it does a great job of demonstrating some of the capabilities of deep learning.
Object Detection Use-Cases and Implications
Some immediate use-cases for this demo could include:
- Counting Pedestrian Traffic
- Targeted Advertising
- Customer Behaviour Analysis
- Actuarial Metrics
- Event Operations
While previously this would have required several computer-vision PHDs with specialised skill for an unbounded amount of time, deep-learning has allowed us to train this targeted model to achieve our desired results requiring only the help of a couple of deep-learning specialists, in a short time-frame with no domain skill dependencies.
If you can think of a similar use-case (which shouldn’t be hard) that your business would benefit from then the solution is closer to your grasp than it has ever been. If you’d like help getting started then Silverpond can get you up and running very quickly.
Don’t hesitate to get in touch!