EcoSort

Inspiration

Hi, my name is Saif along with my teammates Aryan and Shayaan and this is our project to Map Hacks 2 and Open Hacks. When we saw the prompt and prize of sustainability and knew this stat that just in the U.S 35% of the people don’t recycle and sort waste properly, EcoSort was a no-brainer. Our project, EcoSort, can help people around the world whenever traveling or not to put their waste into the proper bin and make our world a better place one piece of waste at a time. Anytime you travel and don’t which bin an item goes in, you can use our website and get instant results of where to put your waste. With ongoing climate change fears and need for action, we decided to jump in and use technology to help make a genuine and significant impact on our community, our world and push our society towards a green future.

What it does

Our application allows people to easily figure out what type of trash it is and where to throw it away in a proper fashion. The user can take an image from either a live camera on the phone or a photo from there file explorer. This is usable on all devices, including smartphones and tablets. After the user takes an image, it will take less than 5 seconds to give a response back which bin to throw the object in.

How we built it

We built the website for using the waste sorter image model with NextJS, React, Vercel. The model itself was a Convolutional Neural Network (CNN) created with TensorFlow, TensorFlowJS, Kaggle for the dataset, and Juptyer Notebooks for the code. We built the model with some data augmentation to make it more accurate as well.

Challenges we ran into

Couple of challenges we ran into was TensorFlowJS had some compatibility issues with TensorFlow when it came to data augmentation, so we have to work around it with some other techniques such as manual rescaling and more. We also ran into issues with exceeding RAM limits on the server making our models due to the amount of data we were processing, so we had to implement data optimization for less RAM usage. Also, we have problems with CORS header while making the website.

Accomplishments that we're proud of

Couple of accomplishment we are proud of is that the model did not overfit, making it really accurate. Also, we are proud that this is running a live demo for anyone to do. Finally, we are proud that is a wide variety of waste it can classify for our users. (We would also like to be applicable for all prizes!)

What we learned

We learned many critical aspects such as the issues of having such a large dataset and with compatibility between TensorFlowJS and TensorFlow. We also learned how to host a model that is accessible outside of the server made by the computer, and how to make a website accessible to a wider range of users, including smartphone and tablet users.

What's next for EcoSort

The next bit to make our dataset to make it more accurate and more classification like glass or metal. Also, we are hoping to make our model more accurate by removing the background for better accuracy.