A comparison of Computer Vision strategies for Lean Construction: experiments, lesson learning and best practices

The construction industry faces major challenges in waste management, where computer vision can support Lean Construction by improving on-site monitoring and material segregation. However, limited guidance exists for selecting appropriate computer vision strategies under real-world conditions. This paper presents a comparative study of three approaches: object detection, binary classification, and anomaly detection, evaluated on a common dataset acquired under operational conditions. Their performance is analysed in relation to accuracy, data requirements, robustness to environmental variability, and the ability to handle unknown materials. The results highlight clear trade-offs between precision, scalability, annotation cost, and sensitivity to changes in acquisition conditions. Based on these findings, the paper proposes a practical decision framework linking algorithmic characteristics with operational constraints such as data availability, environmental stability, and system openness. The main contribution is to bridge the gap between technical performance and deployment feasibility, providing actionable guidelines for practitioners and researchers aiming to implement computer vision solutions for construction waste management.