← Luciad blog

Meet the Luciad hackathon winners: Team Perceptrons

Meet the Luciad hackathon winners: Team Perceptrons

On February 15&16, Luciad organized its first internal hackathon. During 24 hours Luciad employees had the opportunity to develop the project of their dreams. In this series, we’ll zoom in on the winning projects.

Perceptrons Product Diagram Luciad

On the first day of the new year of the dog, team Perceptrons with Runxia Ye, Emeric Beaufays, Ronald Sok and Bruno Rocha Pereira wanted to challenge the saying “You can’t teach an old dog new tricks”. Luciad is turning 20 next year, which could be considered of a certain age by some, so the Perceptrons wanted to find out if Luciad software is still able to learn. If that would be the case, Luciad’s tagline “Connect, visualize, analyze, act” could be changed into to “Connect, LEARN, visualize, analyze, act”.

The team started by showing Lucy, the application template of Luciad’s desktop solution, some basemaps containing the aerial images of the city of Potsdam. Using Lucy's drawing tools, they created some polygons on top of buildings or roads in corresponding classification layers. This were the examples Lucy could learn from using Google's LeNet.

After this learning step, the Perceptrons presented parts of an image for Lucy to provide the most probably classification of that image tile. Lucy replied by adding the appropriate colors associated with the objects.

Meet the Luciad hackathon winners: Team Perceptrons

In the video below you can see a house being annotated with a polygon. This means that everything that is inside the polygon will be considered as “house”. As we zoom out you can see that 3 classes have been annotated; houses, roads and “other”. We then push the “learn” button. This will train a Machine learning model that is able to recognize 32 x 32 images as either being a house, a road or “other”. The model is a convolutional neural network, ideal for image recognition.

Then we hit the “analyze” button which will pick (overlapping) 32x32 images and send them to the neural network which will tell us if the image represents a house, a road or other. The 32X32 blocks are then coloured: yellow for house, green for road and blue for “other”. Finally we compare Lucy’s classification with the real image. 

The Perceptrons didn’t only obtain an accuracy of 80% with the model they built, they also won the first prize of the jury with their project! The team donated their prize, along with a demo of the project, to Coderdojo, an organizer of free programming workshops for girls and boys from 7 to 18 years old.

Meet the Luciad hackathon winners: Team Perceptrons

We hope we convinced the young programmers that some old dogs are still able learn new tricks and who knows that one day some of them would like to work at Luciad...

twitter logo@LUCIADconnect