Posted by Appsilon Appsilon at Hack4Environment
Appsilon has recently taken part in the Hack4Environment – a 24-hour hackathon focusing on dealing with the critical problem of illegal waste disposal (known as fly-tipping).
Read about our AI4Good initiative – Here’s how Machine Learning can be used to analyze wildlife camera trap datasets.
Long story short – in 24 hours, we proposed a solution that granted us an honorable mention in the competition!
Navigate to a section:
About the Hackathon
As said previously, the hackathon’s goal was to somehow deal with the problem of illegal waste disposal. Our solution was based on two parts:
- Application portion – managed by Dominik Krzemiński
- Machine learning portion – managed by Jędrzej Świeżewski
We ended up proposing a solution from three different perspectives, which are discussed in the following section.
Our Solution
Our idea was to build an app addressing the issue from the following perspectives:
- Citizen perspective – citizens could localize sites in their neighborhood with disposed waste and earn recognition for reporting on cleaning it up.
- Local authority’s perspective – we provide them with insights about the waste’s location on a larger scale, allowing them to direct cleaning efforts.
- Manufacturer perspective – we build a mechanism for incentivizing the manufacturers of goods that get thrown out as the waste to act on the information that their products are perceived as littering the local communities.
As icing on the cake, we have built and deployed a Shiny application called Kropla. It aims to display the data provided by the hackathon organizers clustered into localized groups of waste items. We have added a mechanism for claiming waste removal and receiving recognition for it.
In addition, we’ve also trained a computer vision model to detect the brand of a given waste item. The model’s validation accuracy reached 75% on a dataset with 46 classes.
If you have a couple of minutes to spare, please watch our hackathon submission video below. It’s in Polish, so apologies to our non-Polish readers:
Lessons Learned
At the end of the day, there are four main takeaways for us from this hackathon:
- Hackathons are a great occasion to bond within the team. Somehow, a chat late at night serves this purpose best.
- Agglomerative clustering is perfect for dealing with clustering on huge scale differences. The data we dealt with spanned the globe, while it, of course, makes no sense for a fly-tipping site to be of a size larger than a few tens of meters. Many other clustering algorithms (including the common HDBSCAN) struggle with such data leading to, e.g., entire cities becoming a cluster.
- The 75% accuracy reached by our machine learning model does not seem to stand out. When one considers the fact it was an accuracy over 46 classes, it starts feeling very good. However, this is misleading as the data set was heavily imbalanced, with one brand taking up around half of the data. This means a majority-class proxy would have ~50% accuracy regardless of the multitude of classes. Our model’s key strength lies in the fact that it dealt well with some of the tiny classes (for some, the brand was hard to recognize even for a human eye). The takeaway here is that reporting a model’s performance with a single number is typically not sufficient.
- Our team had two participants, while each of the three teams that ranked above us had more (3, 5, and 5) – having more people on board, especially in a limited time setting, can be a differentiating factor.
Conclusion
All in all, the hackathon was a lot of fun! Moreover, it touched a globally and locally important ecological issue and helped raise the awareness of waste reduction and fly-tipping.
We would like to congratulate the other winners, thank the organizers and applaud all the participants for their effort. We regret that we could not gather in person due to the circumstances and mingle with all the people involved in the event to get to know each other beyond the chat communication.
Learn More
- 5 Ways to Apply Ethics to AI
- Fast.ai in R: How to Make a Computer Vision Model within R Environment
- Satellite Image Analysis with fast.ai For Disaster Recovery
- AI for Good: Fighting COVID-19 with Data Science
- AI for Assisting in Natural Disaster Relief Efforts: The xView2 Competition
Appsilon is hiring for remote roles! See our Careers page for all open positions, including R Shiny Developers, Fullstack Engineers, Frontend Engineers, a Senior Infrastructure Engineer, and a Community Manager. Join Appsilon and work on groundbreaking projects with the world’s most influential Fortune 500 companies.
