Current evaluation paradigms in Document Layout Analysis (DLA) overwhelmingly focus on measuring the quality of detected elements through metrics such as Intersection over Union (IoU), mean Average Precision (mAP), and F1-score. While these metrics assess detection accuracy, they remain structurally silent on a critical question: what has been missed? In industrial contexts where documents feed safety-critical processes—energy infrastructure maintenance, pharmaceutical compliance, financial auditing, legal contract analysis—an omitted layout element can propagate significant downstream consequences that far exceed the cost of a misclassification. This paper argues that the field requires a fundamental shift from detection performance to completeness certification. We introduce the Completeness Confidence Index (CCI), a conceptual framework that aggregates three independent proof vectors—residual signal analysis, structural coherence validation, and cross-modal redundancy—to estimate the probability that a layout analysis has captured all semantically relevant regions of a document. We formalize the notion of informative void, drawing on epistemic uncertainty quantification, conformal prediction theory, and probabilistic document grammars. Rather than presenting experimental results, this position paper establishes the theoretical foundations and formalizes the research agenda, calling for the creation of an “Omission Challenge” benchmark and for process-dependent calibration of completeness metrics. We argue that as AI-driven document analysis becomes pervasive in industrial pipelines, neutralizing uncertainty about what remains undetected is not merely an academic concern but an operational imperative.
Gabriel Zo-Hasina Rasatavohary (Thu,) studied this question.