Waste object detection with subtype-aware fusion and efficient re-ranking (SAFE-R) for hazardous waste identification and recycling in waste streams

Abstract

Waste management is essential for a multitude of reasons, including protecting the environment and the people who are employed for this job, especially in developing countries like Bangladesh where primitive practices of waste-collection and landfilling still persist. Improper classification and disposal of waste can lead to many issues like air and water pollution. This study aims to address the challenges of waste management by proposing an automated system that classifies waste into seven different categories like sharpened objects, plastic, metal, glass, etc., using computer vision. To that end, multiple modifications of the popular YOLOv8 model were tested to ascertain the ideal balance between accuracy and latency for real-world implementation. Therefore, after filtering the set of 2,444 images (24,715 annotated objects) and trying various architectural variations, we discovered that the YOLOv8m detector has the balance of being the fastest and most accurate (mAP50 = 0.591, mAP50-95 = 0.332, around 9 ms per image).Furthermore, our study aims to identify hazardous objects like syringes amongst medical waste and nails amongst sharpened objects using query-aware reranking. When identifying dangerous subclasses our pipeline, too, showed significant improvement compared to an open-vocabulary baseline (OWLViT). To illustrate, with an IoU of 0.5, syringe detection had an AP50 of 0.619 versus 0.064 with OWL-ViT and nails had 0.538 versus 0.214. These findings indicate that we can successfully detect the presence of dangerous objects such as syringes in real time, which makes waste collection not only quicker and more efficient but also significantly safer to the people working on it. Implementation of proper waste classification using our proposed pipeline can give the waste-collectors safety and promote practices like composting and recycling. In this thesis, we propose to introduce an automated system that will help classify and separate garbage into superclasses and further distinguish subclasses by the combination of YOLOv8m and CLIP algorithms which can aid garbage collectors in their task and foster better recycling and landfill practices.