Make Pets Found (using AI) - Day 3

3 minute read

Day 3

Climbing out of bed today and there was no denying it: we were definitely not in Bristol. It was an early November morning, the sun was beating down, and there was a herd of goats by the swimming pool. Fortunately they weren’t unsupervised, although Henry was ready and willing to try his hand as a goatherd.

George, Sam and Henry on laptops beside pool

Much of the day was spent experimenting with deep learning, and how we can manipulate images to get better results. The main aim was to get an end to end example of object detection working using the AWS Object Detection algorithm - so for this we were experimenting with the Caltech birds dataset because it was already well formatted for what we needed it for, including species and bounding box data. 

It’s worth noting that Machine Learning is very much within the Python developer’s comfort zone, and as the majority of the developers working on this side of the project are expert in Javascript, today was a real learning curve. A challenge, but one that we embraced fully.

Henry, James and Simon on laptops outside

Sam, George and Henry ploughed on with Python and managed to get an end-to-end example up and running which manipulates a dataset, trains a machine learning instance on AWS SageMaker and deploys an endpoint. The deployed endpoint can then be hit with an image uploaded to our app, processed by the trained model, and will predict the likely species and bounding box dimensions for where the bird lies in the image in JSON, so that we can use it in our JS-based applications. Now that we’ve got a process working with birds, we’ll move on to cats and dogs.

Meanwhile, the rest of the devs focused on getting the app set up to integrate with Google Cloud Vision. We are expecting to rely on this API quite heavily when it comes to identifying traits like breed, coat pattern and age of lost pets so that we can search using these features in the final app.

James working on laptop

At the end of the day James presented to the team some of the awesome work he’d done on the frontend app, in what was a masterclass of ‘how to run a demo smoothly’.  

The sun still shining, beer o’clock was announced with the snap of a San Miguel ring pull. As Matt stalked the poolside area taking pictures of various creatures (cats, critters, Gravitywell devs), James cracked on with dinner, demonstrating to a doubting George that vegan food can actually be prettyy, prettyy, pretty good.

Hugo with a San Miguel beside pool

View next (Day 4)

View previous (Day 2)

If you’d like to learn more about how Gravitywell can help you benefit from a hackathon, then please get in touch.