Iman Ranjbar, Yiannis Ventikos, Mehrdad Arashpour
The construction and demolition (C&D) sector is a major contributor to Australia’s total waste, with reports indicating continuous annual growth. Among this waste, end-of-life plastic represents a valuable recyclable material with significant recovery potential. However, effective separation of plastic waste remains a challenge due to the highly cluttered and heterogeneous nature of C&D waste. This study focuses on the detection and segmentation of C&D plastic waste using advanced instance segmentation models to enable efficient waste sorting and recycling. A large, specialised dataset is curated, capturing the complex and very often deformed nature of C&D plastic waste across seven key categories: buckets, cables, drums, insulation, liquid containers, pipes, and PVC profiles. The dataset features a high density of objects per image, ensuring robust model training and generalisation in real-world scenarios. State-of-the-art instance segmentation models, including FastInst, RTMDet-Ins, YOLOv9, and YOLOv11, are trained and evaluated on this dataset. Among these, YOLOv11 demonstrated the highest performance, achieving a mean Average Precision (mAP) of 51.3 while maintaining a real-time inference speed of 89 frames per second (FPS). A systematic analysis of the models’ strengths and limitations is provided, highlighting challenges associated with segmenting highly cluttered and overlapping objects. Additionally, EigenCAM visualisations are used to interpret the model’s decision-making process. The findings demonstrate that the proposed models achieve accurate segmentation of C&D plastic waste, contributing to improved resource recovery and the advancement of a circular economy within the construction industry.