TreasureBot 3000
It is hard, as a foreigner moving to Germany, to not be struck by how seriously germans take recycling seriously. We all know the 4 horsemen of households that got us all confused at least once: blaue Tonne, Wertstofftonne, Biomüll and Restmüll. In fact, the number of categories of waste is up to 12. The idea behind TreasureBot 3000 was therefore to create a tool able to help a user to know in which bin a waste should be disposed. The project has been developed by a team of 4 people as part of the 2-week project of Le Wagon coding bootcamp.
The core of the app relies on the usage of the "ResNet50" pre-trained model, to which a single dense layer has been added together with a 10 neurons predictive layer corresponding to our 10 output categories:
Paper and Cardboard
Plastic (plastic containers, wrappers, straws…)
Aluminium (food cans, aluminium foils…)
Glass (except bottles)
Drinking cans
Plastic bottles
Glass bottles
Organic
Clothes and shoes
Batteries
The model has been trained on approximatively 700 pictures for each class mainly taken from Kaggle. As a result, the app tells the user in which bin the predicted waste should go, and suggests to look for a Pfand on plastic and glass bottles as well as drinking cans before throwing them away. For the items that cannot be disposed in one of the 4 regular bins, another feature of the app includes a map to find the closest bin for glass, clothes and batteries as well as Pfand return machines. The API for this application has been made with FastAPI and is currently running on GCP. The web interface has been created with Streamlit.