Automating construction and demolition waste (CDW) sorting remains difficult because materials vary widely in shape, scale, appearance, and often occur in heavy clutter. Existing datasets fall short of capturing this complexity, which limits the reliability of current computer vision models. This study presents a benchmark dataset of 408 high-resolution CDW images containing nearly 10,000 annotated objects across seven material categories (brick, concrete, glass, metal, paper, plastic, and wood). Images were collected from skip bins, active construction and demolition sites, and municipal solid waste bins to reflect the clutter and disorder typical of primary sorting conditions in material recovery facilities (MRFs). Using this dataset, we evaluate one end-to-end instance segmentation model and two modular pipelines that pair state-of-the-art detectors with a foundation segmentation model, enabling clear diagnosis of model behavior and simple future upgrades. Transformer- and Convolutional Neural Network (CNN)-based detectors show distinct scale-dependent strengths. A key operational insight is also identified. Increasing the confidence threshold from 0.01 to 0.1 in the Transformer-based modular pipeline reduces inference latency by 63.81% with only a 1.42% drop in mean average precision. This positions confidence tuning as a practical mechanism for managing the trade-off between throughput and material purity. Overall, this work introduces a realistic CDW benchmark dataset, establishes modular recognition pipelines, and offers actionable guidance for deploying scalable, high-performance sorting systems in MRFs.
Malla et al. (Fri,) studied this question.