Contaminated Identification

This project, which started five years ago in Boston, combines an online platform, a data collection system, and a mobile app to analyze the content of bins based on visible elements. It is the team's longest-running project, beginning in 2021, and is led by a young startup of around 10 people. Spare-it has been a key partner of Boston University, where IoT devices and sensors deployed in the bins throughout the CDS building send real-time data every second to our system, which is then parsed and displayed on our admin dashboard for analysis. Additionally, students take weekly photographs of bin contents using the mobile application, and these images are manually labeled and added to the dataset daily. This combination of automated data collection and manual labeling allows for continuous improvements in waste analysis and reporting. Spare-it is now expanding the product's potential by integrating modern AI and machine learning techniques, focusing on neural networks for generating image segmentation models. During the Spring 2024 and summer terms, Spark! students developed a basic machine learning pipeline using the YOLOv8 framework by Ultralytics. Synthetic images were introduced to augment the dataset, though results were mixed. The model's performance varied significantly across 101 classes, with selected classes achieving around 68% accuracy overall. This semester, students will aim to further improve the model’s accuracy, building upon the insights from previous work. A key takeaway from earlier efforts is that enhancing data quality is crucial. The focus this semester will be on refining the accuracy of individual objects within the dataset, identifying the best labels, and fine-tuning the model to boost the precision of object detection and classification. Below are the key tasks as suggested by the client. Students do not need to complete all of them but should work with their team members, PM, Spark! staff, and instructor to come up with a plan and milestones for the semester. The goal of this project is to improve the existing contamination identification model, whether through dataset enhancement or model improvement.