Beyond the Benchmark: An Iterative Circular Economy Data Design Workflow for Post-Consumer Plastic Detection in Light-Packaging Sorting Facilities

A growing body of public waste datasets and increasingly capable object-detection algorithms suggests that waste detection may be approaching a solved computer-vision problem. Operational lightweight-packaging sorting, however, represents a different problem class. The underlying waste stream is highly heterogeneous and only partially characterized, relevant objects occur unpredictably, and the appropriate representation of the detection task cannot necessarily be specified completely in advance. In such settings, the challenge is not only to improve a given dataset, but to determine iteratively what should be observed and how observations should be represented for the intended operational task. This paper investigates how operational context, annotation experience, and model evaluation can inform interventions in data generation and representation for detecting post-consumer plastic objects with potential for selective recovery. Using Design Science Research Methodology, we develop an iterative data-design workflow through empirical activities conducted within the RecycleBot project at a German lightweight-packaging sorting facility. Across three field campaigns, four collection strategies modified how physical objects were selected, presented, and observed: passive collection, object reintroduction, purposive object enrichment, and controlled multiview capture. Model-informed evaluation, complemented by contextual knowledge, supported the examination of prediction–annotation disagreements and potential limitations in data generation and representation. The annotation process further demonstrated that physical objects and detection instances do not necessarily correspond one-to-one, requiring explicit, task-informed representation decisions and transparent annotation conventions. The resulting contribution is an empirically grounded, iterative data-design workflow that connects data evaluation and diagnosis with targeted interventions in data generation and representation. Rather than prescribing an optimal collection strategy or treating reference annotations as fixed ground truth, the workflow makes data-design decisions explicit and open to revision in response to identified limitations and ambiguities. Its contribution lies in structuring task- and context-informed interventions within a partially characterized operational environment, while its prospective effectiveness remains to be validated.