Gaia's League

Gaia's League aim at cleaning the planet. Take a picture of a dirty place, clean it, then take another picture of the result. Our ML algorithm will count the garbage collected and reward with point.

Comment

Inspiration

As we like to spend time in park running, we remarked an increase in the number of garbage in park, forest, etc... Unfortunately, no one like cleaning place cause it is dirty, time-consuming but especially not fun at all ! We tried to create an application to encourage children and their parents to clean the planet by making it fun !

What it does

First you need to find a dirty place in a park for example. You start the application and take a first picture. You then clean the place as fast as you can and when you have finished, you take another picture of the cleaned place.

Our machine learning algorithm will compare the two pictures, count the number of garbage in each picture respectively, and reward you with points accordingly to the number of garbage you have collected in a limited time.

Our algorithm also compare the two picture and check that the place is the same one (in order to avoid cheating).

How we built it

We used React Native for the application part. The pictures are hosted with Firebase / Firestore and their are send to a rest API written in Python / Django. The neural network is built with Pytorch. To be more precise, we used a efficientDet architecture (the state-of-the-art in terms of image segmentation) with transfer learning already trained on ImageNet.

Challenges we ran into

The design of the application and also the UI. As our main target is children we had to think about designing something that can be fun / easily design for children ! The neural network training and tweaking was also time-consuming, however we found good database for garbage detection.

Accomplishments that we're proud of

We succeeded to create the application quite fast and release it on the Google Play Store without major bugs. We also are quite proud of the neural network construction and the fact that it works correctly !

What we learned

To be more confident with React Native and Pytorch. Also, it was the first time that we were combining a neural network with a mobile application, so we have learned how to publish the API and try to optimize it to give results as fast as possible to the application.

What's next for Gaia's League

Optimization of the algorithm (the algorithm takes ~10 seconds to detect both images and to upload), we probably can faster this by using a better image compression and a better hosted machine.

Share it with many users and get feedback to continue to work on :) We also think about using for garbage detection in see for example.

Share this project:

Updates