A detector report may display one prominent score. Its meaning depends on which categories, matching rules and evaluation images produced it.
Ultralytics' metrics guidance describes precision-recall-based detection measures and evaluation under overlap thresholds. Ultralytics — metrics.
Request the breakdown
For a multi-detail character, inspect category-level results as well as the aggregate. A strong result on easy body detections can coexist with weak accessory detection. Identify the actual categories that matter to the intended review.
Compare equivalent evaluations
Do not compare two model scores until dataset boundaries and evaluation settings align. A result from unrelated public images does not establish performance on a supplier's photographs. Retain failed examples alongside the summary so the team can judge practical consequences.
Source reviewed 25 October 2026. Interpretation and original diagrams: PLUSHIP / Qunhe. Examples are illustrative.
Related reading
Adaptive Thresholding Responds to Local Image Conditions
Object Detection Can Locate a Detail Without Measuring Its Quality
Combining Models Helps Only Under the Actual Evaluation Conditions