Calculator Inputs
Enter ranked detections for every class and optional IoU slice.
Formula Used
Predictions are sorted by confidence before cumulative counts are calculated.
Eleven-point and one-hundred-one-point methods sample fixed recall levels. Trapezoidal mode integrates neighboring raw precision-recall points.
How to Use This Calculator
- Enter one class or IoU slice on each line.
- Provide the total ground-truth objects after the first pipe.
- Add confidence scores with matched TP or FP labels.
- Select an AP method and preferred headline average.
- Set confidence filtering and maximum detection limits.
- Press the calculation button to display results above.
- Review curves, class scores, false positives, and missed objects.
Mean Average Precision Insights
Understanding Mean Average Precision
Mean average precision evaluates ranked predictions across several detection classes accurately. It combines precision and recall into one comparable summary score effectively. Each class receives average precision before class scores are finally combined.
Why Ranking Quality Matters
Detection systems usually return many predictions with calibrated confidence scores. Correct objects should appear before incorrect or duplicate detections consistently. A strong ranking preserves useful precision while recall steadily increases reliably.
Precision and Recall
Precision measures how many accepted detections are actually correct overall. Recall measures how many known objects the model successfully finds. Both measures change whenever another ranked prediction is evaluated sequentially.
Average Precision Meaning
Average precision summarizes the complete precision-recall curve for one class. Interpolation reduces small precision fluctuations caused by ranking transitions smoothly. Different standards use different recall points and integration rules.
Mean Average Precision
Mean average precision averages class results into one headline metric. Macro averaging gives every evaluated class equal influence consistently. Weighted averaging gives larger ground-truth classes more influence proportionally.
Understanding IoU Labels
Object detection matches predictions using intersection over union thresholds. Higher thresholds demand tighter overlap between boxes and targets. Your input labels must already reflect the selected threshold accurately.
Reading the Plot
The horizontal axis shows recall from zero through one. The vertical axis shows precision across the same bounded range. Better curves remain high while extending farther toward full recall consistently.
Interpreting Class Differences
Classwise curves reveal categories hidden by one overall score. Low precision may indicate confusion, duplicates, or poor confidence ranking. Low recall often indicates missed objects or aggressive filtering.
Threshold Effects
A confidence threshold removes predictions before evaluation and plotting. Raising it can improve precision while reducing attainable recall. Lowering it includes more candidates but may introduce extra errors.
Maximum Detection Limits
Detection limits reflect deployment rules or benchmark evaluation settings. Small limits may discard useful lower-ranked true positive predictions. Large limits expose more false positives and ranking weaknesses clearly.
Choosing an AP Method
Continuous interpolation suits many modern object detection evaluation workflows. Eleven-point interpolation supports comparisons with older benchmark conventions. One-hundred-one-point interpolation closely follows common modern reporting practices.
Using Results Responsibly
Compare models only when datasets, matching rules, and thresholds match. Review class counts because tiny classes can produce unstable scores. Always inspect curves before trusting any single aggregate metric.
Improving Model Performance
Investigate false positives near the top of each ranking. Check missed objects across sizes, scenes, and difficult conditions. Better labels and calibration often improve evaluation consistency.
Practical Reporting
Report the AP method beside every published mean score. State confidence filters, detection limits, and matching thresholds clearly. Include classwise results so important failures remain visible clearly.
Frequently Asked Questions
1. What does mean average precision measure?
It measures ranked prediction quality across evaluated classes. Each class receives an average precision score. Those scores are then combined using the selected averaging rule.
2. What does a TP label mean?
A TP label marks one correct matched detection. The prediction should match an unused ground-truth object. Duplicate matches should use an FP label.
3. Can I evaluate several IoU thresholds?
Yes. Add separate lines using names like Cat@0.50 and Cat@0.75. The calculator reports every slice and averages their scores.
4. Which AP method should I choose?
Use the method required by your benchmark. All-points interpolation suits many modern workflows. Fixed-point methods support specific historical or standardized comparisons.
5. Why can confidence filtering lower recall?
Filtering removes low-scoring predictions before evaluation. Some removed predictions may be true positives. Their removal limits the highest reachable recall.
6. What is macro mAP?
Macro mAP gives every class or evaluation slice equal weight. Large and small classes influence the result equally. This exposes weak minority classes more clearly.
7. What is weighted mAP?
Weighted mAP uses ground-truth counts as class weights. Larger classes influence the final score more. Review classwise AP alongside this summary.
8. What is micro AP?
Micro AP pools retained detections across all evaluation slices. It sorts them together by confidence. The pooled ranking then produces one precision-recall curve.
9. Why does the plot show an envelope?
The envelope replaces each precision with the best later precision. This creates a monotonic interpolated curve. Disable that option to inspect raw ranking transitions.