EcoSortAR

Inspiration

The problem we are tackling is recycling contamination. This is when waste items are not sorted into the correct bins. Canadians produce 36.5 million tonnes of solid waste annually, and only 27% of that gets properly diverted to recycling. Across North America, this value is around 25%, meaning items that should’ve been recycled get thrown out in the garbage. As a case study, in Toronto, this recycling contamination adds up more than 20 million in extra recycling costs. This is a multi billion dollar problem, in the US recycling contamination adds more then 3.5 billion in unnecessary costs through manual sorting labour, reduced recycling material value, and landfilling rejected loads.

What it does

Our solution is EcoSortAR - its an app that uses machine learning to instantly identify what category your waste item belongs in and uses augmented reality to tell you what it is. It analyzes your camera feed in real time to determine what type of item you have to dispose of, so that over time, you’ll learn how to sort correctly without having to guess.

How we built it

For the backend, it uses a YOLO real time deep learning model for object detection and was trained on a dataset of 1000 images of different items sorted into categories like e-waste, organic, plastic, cardboard, and paper

Challenges we ran into

One of our main challenges was finding a suitable dataset that matched our problem, since most available datasets were limited or didn’t reflect local waste conditions. We also faced imbalanced classes, as some types of waste like trash had far fewer images, making the model less accurate on them. Another challenge was the visual similarity between materials such as cardboard and paper or plastic and glass, which often confused the model, along with the real-world variability of garbage, where lighting, angles, and messy or crumpled items looked very different from the clean dataset images.

What's next for EcoSortAR

For the future, the goal is to reduce sorting errors by 10-20%, tailor sorting categories to each user’s geographic location, and gamify correct recycling by rewarding consistent accuracy through a points system. Another application would be to integrate a more portable version of this into meta glasses and other AR devices for more convenience.