Evaluation Protocols and Benchmark Interpretation for Deep Learning-Based Object Detection: A Scale-Centered Review

This is the authors' abstract. We don't add key points for this paper.

Object-detection benchmark scores are often transferred across settings without restating the evaluation target that gives a metric its meaning. This focused interpretive review examines how that target changes along a scale-centered path from regular-scale to small-object detection, with point-based targets treated as a representation-dependent branch. Orientation-aware and temporal evaluation are treated as selected extensions when geometric representation or sequence-level evaluation targets require criteria beyond scale alone. Across these settings, the review distinguishes reporting protocols, matching and localization measures, and loss-derived formulations used mainly for optimization. An intersection-over-union (IoU) threshold can determine whether a prediction is matched to a target, while precision and recall describe an operating point and average precision, mean average precision, and average recall summarize behavior under a specified benchmark protocol. The same metric label can therefore support different claims when thresholds, annotation granularity, prediction structure, or temporal targets differ. A four-slot evaluation contract makes the prediction target, matching primitive, reporting endpoint, and metric role explicit. A cross-setting guide links those slots to protocol selection. Reassessments of Common Objects in Context (COCO), Dataset for Object Detection in Aerial Images (DOTA), and multi-object tracking (MOT) predictions connect scale-conditioned reporting, geometric matching, and identity continuity to concrete interpretation decisions. The review synthesizes these findings into a practical guide for reading benchmark evidence and choosing evaluation protocols.