Efficient Waste Sorting for Circular Economy: A Confidence-guided comparison between One-Vs-All and One-Vs-Rest Classification Strategies with Human-in-the-Loop for Automated Waste Sorting

Effective household waste sorting is a critical prerequisite for a functional Circular Economy, yet it remains hindered by complex and heterogeneous municipal regulations. AI-based mobile applications offer a promising solution to guide residents. However, these systems must be both accurate and easily adaptable to local sorting schemes. This paper evaluates two primary classification strategies—One-Vs-All (OvA) and One-Vs-Rest (OvR)—to determine their performance and their ability to identify uncertain predictions for human-in-the-loop refinement.

The researchers constructed a six-class waste management dataset based on the sorting scheme of Goslar, Germany, incorporating images from the TACO and TrashNet datasets to improve representativeness. The OvA approach uses a single multi-class classifier with a softmax output, while the OvR approach employs six independent binary classifiers, each using a sigmoid output to distinguish one class from all others. The study systematically compares these strategies by applying confidence thresholds to categorize predictions into four groups, ranging from high-confidence to no-vote samples, to assess the trade-off between classification accuracy and the human effort required for data annotation.

While the OvA strategy achieved marginally higher overall accuracy, the OvR strategy proved superior for practical, iterative model improvement. By analyzing the confidence levels of the binary classifiers, the researchers found that the OvR framework effectively isolates ambiguous samples. Specifically, by focusing human review on low-confidence groups (multi-vote and no-vote samples), the OvR strategy allows for the correction of more than 50% of all misclassifications while requiring human intervention for less than 5% of the total dataset. In contrast, the OvA strategy lacks this granular uncertainty detection, making it harder to identify which samples require expert review.