MDWD: A Street-Level Dataset for Municipal Solid Waste Detection in Dense Urban Environments
This paper introduces the Maltese Domestic Waste Dataset (MDWD), a comprehensive street-level benchmark featuring 3,697 high-resolution images and over 11,000 annotated instances of domestic waste, to address the gap in municipal solid waste detection resources and establish reproducible performance baselines for various object detection models.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints, you are looking for trash. This is the world of Computer Vision, a branch of science where we teach computers to "see" and understand the world through cameras, just like humans do. In this field, a key concept is Object Detection, which is like teaching a computer to not only spot a cat in a photo but also draw a box around it and say, "That is a cat!" Another important idea is Machine Learning, where the computer learns by practicing on thousands of examples until it gets really good at the job. Why does anyone care? Because our cities are getting crowded, and keeping them clean is a huge, messy challenge. If we can teach computers to automatically spot piles of garbage on the street, city workers could clean up faster and smarter, turning a chaotic, reactive job into a smooth, automated operation.
Now, meet the MDWD, a new "training gym" for these computer detectives. Before this paper, most computer programs trained on trash had it easy: they looked at single pieces of litter on a clean table, or they looked at trash from high up in the sky like a bird. But real life is messy. Trash is often hidden behind other objects, squished into bags, or sitting in confusing piles on crowded city streets. The authors from the University of Malta realized that to build a truly smart trash-spotting robot, they needed a dataset that looked exactly like the real, chaotic streets of Malta. So, they created the Maltese Domestic Waste Dataset (MDWD).
This dataset is a massive collection of 3,697 high-resolution photos taken right at street level. Inside these photos, the team manually drew 11,461 boxes around different types of waste, teaching the computer to recognize five specific categories: Mixed Waste (black bags of leftover household trash), Organic Waste (small white bags of food scraps), Recyclable Material (gray or green bags with paper and plastic), Orange CMD (distinctive orange bags from the city's cleaning crew), and Other Waste (the weird, rare stuff like big furniture or glass bottles). The photos capture all the difficulties of the real world: shadows, rain, objects blocking each other, and trash of all different sizes.
To see if this new dataset actually works, the researchers put it through a rigorous test, acting like a coach running a marathon for different types of computer brains. They pitted several generations of YOLO (a popular family of fast, real-time detectors) against a newer, transformer-based model called RF-DETR. Think of YOLO as a sprinter who is fast and good at spotting things quickly, while RF-DETR is like a careful observer who looks at the whole picture to understand context.
The results were clear: the RF-DETR-M model was the star of the show. It achieved a 94.49% success rate (measured as mAP50) and an F1-score of 93.56% on the test set, meaning it was incredibly accurate at finding and categorizing the trash. However, the paper also found that smaller, lighter versions of the models (like the "Nano" and "Small" variants) were still very competitive, proving that you don't always need a giant, heavy computer to do a good job. The researchers explicitly noted that while the Orange CMD bags were easy to spot because of their bright color, the "Other Waste" category was the hardest, often missed because it contains such a weird mix of rare items.
Crucially, the paper does not claim to have "solved" the problem of trash forever. Instead, it suggests that the MDWD is a solid, reproducible foundation. It proves that street-level, instance-level detection is possible and provides a standard "ruler" for future researchers to measure their own progress against. The authors argue that this dataset fills a specific gap: no other resource offers street-level photos with detailed boxes around specific types of municipal waste streams. By making this data public, they hope to spark more research into how we can use vision to keep our cities clean, moving from manual inspections to smart, automated monitoring.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.