Principle:Roboflow Rf detr COCO Evaluation
| Knowledge Sources | |
|---|---|
| Domains | Object_Detection, Evaluation |
| Last Updated | 2026-02-08 15:00 GMT |
Overview
The standard evaluation protocol for measuring object detection model performance using COCO metrics (mAP, precision, recall, F1).
Description
COCO evaluation measures detection quality across multiple dimensions:
- mAP@50:95: Mean Average Precision averaged over IoU thresholds from 0.50 to 0.95 in steps of 0.05. This is the primary metric for detection quality.
- mAP@50: Mean Average Precision at IoU threshold 0.50 (more lenient)
- Per-class metrics: RF-DETR extends standard COCO evaluation with per-class precision, recall, and F1 scores computed by sweeping confidence thresholds to maximize macro-F1
The evaluation process involves running the model on validation data, matching predictions to ground truth using the COCO evaluation protocol, and computing aggregate statistics.
Usage
Use this principle to assess model performance after each training epoch or to compare models. The mAP@50:95 metric is the standard benchmark for detection quality.
Theoretical Basis
The COCO evaluation protocol computes Average Precision (AP) by:
- Ranking all detections by confidence score
- Computing precision and recall at each detection threshold
- Computing AP as the area under the precision-recall curve (with interpolation)
- Averaging across IoU thresholds (for mAP@50:95) and classes
RF-DETR's coco_extended_metrics additionally sweeps confidence thresholds to find the operating point that maximizes macro-F1, providing practical deployment metrics.