The paper addresses the challenge of multi-label scene classification, where natural scenes can contain multiple objects and thus be described by multiple class labels. Unlike classic pattern recognition problems where classes are mutually exclusive, multi-label classification involves classes that overlap by definition. The authors propose a framework to handle such problems and apply it to semantic scene classification. They discuss training and testing approaches, introduce new metrics for evaluating individual examples, class recall, precision, and overall accuracy, and demonstrate the effectiveness of their methods through experiments. The key contributions include a novel training strategy called "cross-training," which uses multi-label data multiple times during training, and three classification criteria for testing: P-Criterion, T-Criterion, and C-Criterion. The C-Criterion, which uses the MAP principle to select a threshold, is shown to be effective for multi-label classification. Additionally, two new evaluation metrics, z-Evaluation and Base-class Evaluation, are proposed to assess the performance of multi-label classification systems. The experiments show that the proposed methods are suitable for scene classification and can generalize to other classification problems.The paper addresses the challenge of multi-label scene classification, where natural scenes can contain multiple objects and thus be described by multiple class labels. Unlike classic pattern recognition problems where classes are mutually exclusive, multi-label classification involves classes that overlap by definition. The authors propose a framework to handle such problems and apply it to semantic scene classification. They discuss training and testing approaches, introduce new metrics for evaluating individual examples, class recall, precision, and overall accuracy, and demonstrate the effectiveness of their methods through experiments. The key contributions include a novel training strategy called "cross-training," which uses multi-label data multiple times during training, and three classification criteria for testing: P-Criterion, T-Criterion, and C-Criterion. The C-Criterion, which uses the MAP principle to select a threshold, is shown to be effective for multi-label classification. Additionally, two new evaluation metrics, z-Evaluation and Base-class Evaluation, are proposed to assess the performance of multi-label classification systems. The experiments show that the proposed methods are suitable for scene classification and can generalize to other classification problems.