When traveling abroad, we noticed how complicated it might be to sort trashes correctly, according to the local rules. In Switzerland, trash recycling system can be particularly puzzling for foreigners. We wanted to address that problem.
What it does
Our application helps a user to know in which container should the trash go. What should I do with this aluminum can ? Just upload a picture on our website, and we'll show you examples of appropriate bins.
How we built it
We built a custom dataset, expressing material in portions of the image. Then we trained a CNN using YOLO and Darknet to recognize those specific classes (PET, aluminum, cardboard,...). We obtained a decent accuracy. Then the category is forwarded to the frontend, where the adequate bins are presented to the user.
Challenges we ran into
A major challenge was the absence of existing labelled dataset. Therefore, we created our own to train the model. Training the model was also a challenge, as it was our first time with YOLO and Darknet.
Accomplishments that we're proud of
With this project we had a first contact with YOLO and Darknet. We are proud to add this discovery to our panel.
What we learned
We got familiar with YOLO and Darknet. We learned how to deal with the whole procedure for training a model, ie. from data collection and labelling to model training.
What's next for EasyRecycle
Improve the dataset in terms of amount of labelled pictures. With a bigger amount of labeled trash pictures, we would then be able to train our model much more efficiently. The long-term vision would be to include other countries, with their specific recycling rules. Another scale of interest is the local variation of country-level rules. With finer localization and better knowledge of local particularities, we would then be able to better suit everyone's needs. For computers, IP address could be the most we can get, while with smartphones, GPS can yield a very precise localization.