Wonderful things are happening at Intel’s Ethernet Products Group (EPG) Software Development. EPG software engineers frequently use their technical superpowers to contribute to local communities, education, ecology, and more. And all this after working hours!
Below, discover how Marta Plantykow, a software engineer with the EPG software Drivers team, helped create Detectwaste; an open-source AI-based solution to detect litter.
What is Detectwaste?
Detect waste is a non-profit, educational eco-project that uses artificial intelligence to fight the world’s waste pollution problem. In five months, a team of nine women supported by five mentors and seven partners created a neural network to detect and categorize litter with an accuracy close to 75%.
What motivated you to join this initiative?
This initiative allowed me to do something good for the planet, grow my skill set, meet motivating people, and spend invaluable (in terms of the new-COVID-19 reality) time with them.
How did you develop Detectwaste?
We would meet online every other week to report our activities status, discuss the next steps and our goals. We spent the first sprint on data collection and analysis (with data quality assessment). Then spent a few sprints designing models, training, and testing them.
After we determined the best model, we decided to expand our database; the original consisted of only the Trash Annotations in Context (TACO) database – an open image dataset of waste in the wild. Created by one of our mentors, it consisted of around 5,000 data points and only about 1,500 of them annotated. The rest of the pictures were annotated by one of our partners, epinote.ai. At that stage, we decided to search for other litter-related databases such as TrashCan or TrashNet. Eventually, we grew our litter database to 100,000 pictures and decided to research litter detection.
At the same time, we worked on final adjustments like augmentation methods. Close to the end, we noticed our models worked drastically better when classifying only a single class. Our last pipeline detects litter with the EfficientDet model and classifies detected waste into seven categories with EfficientNet.
What categories of trash does Detectwaste recognize?
Gdansk’s rules inspire waste categories in Detectwaste. Currently, they are classified into five distinctive categories: paper (clean or dry), glass, metal and plastic, biodegradable (without animal waste, fat, or bones), and residual (non-recyclable) waste. You can find more detailed information on our website. Thanks to a variety of data in our datasets (pictures taken at home, in the office, outside in a natural environment, on the street, and even underwater), advanced augmentation methods that detect waste can be used everywhere. It is crucial that the solution can accurately distinguish between different environments.
What are some challenges or key takeaways from this project?
I was surprised by some of the segregation rules. At my home, I use six different bins and pay a lot of attention to details, such as having biodegradable garbage waste bags made of paper for bio litter, but it appeared that I was wrong with some of my actions. For example, I wasn’t aware that chicken bones should end up in residual waste or that a chips bag shouldn’t be a part of metals and plastic, as it is greasy. In terms of challenges—having a great team made development smooth.
What are the next steps?
We want to spread the news about our project and research. Cornell University published it on their page and we are working on more publications. Also, we won’t stop here as we have many more ideas to explore. In October, we are launching our new project; working with sign language to help people who are deaf or hard of hearing.
If you’re interested in additional details or want to join our team, visit our webpage.
Want to learn more?
Detectwaste Home: https://detectwaste.ml/
Detectwaste GitHub: https://github.com/wimlds-trojmiasto/detect-waste