Analysis results
Automatic interpretation
Threshold performance
Comparison test
Detailed IoU records
Data and calculation settings
Formula used
Intersection over Union: IoU = intersection area ÷ union area.
For bounding boxes, the intersection is the shared rectangle. The union equals both box areas minus their intersection.
For binary masks, intersection counts shared positive pixels. Union counts pixels positive in either mask.
How to use
- Choose manual values, bounding boxes, masks, or imported data.
- Enter records and select threshold, filters, grouping, and chart options.
- Press calculate to validate data and create the distribution.
- Review summary statistics, threshold performance, outliers, and class results.
- Export the table, chart, JSON data, or PDF report.
Supported CSV example
sample_id,image_id,class,confidence,iou,model,dataset,split,object_area,match_status img_001,img_001,cat,0.94,0.82,Detector A,Wildlife,test,1840,matched img_002,img_002,dog,0.88,0.63,Detector A,Wildlife,test,9700,matched img_003,img_003,bird,0.51,0.00,Detector A,Wildlife,test,640,unmatched
Interpretation guidance
A distribution near one indicates strong overlap. A concentration near zero indicates missed, misplaced, or poorly sized predictions.
Threshold choices depend on the task. Object detection evaluations often compare performance across several thresholds.
IoU can disadvantage small objects because minor coordinate errors produce large proportional changes. Review size groups before drawing conclusions.
IoU measures overlap, not confidence calibration or class correctness. Combine it with precision, recall, calibration, and error analysis.
Frequently asked questions
What is a good IoU score?
Higher values indicate better overlap. The acceptable threshold depends on the task, dataset, and evaluation protocol.
Can this calculator use segmentation masks?
Yes. Enter aligned binary masks containing zeros and ones.
What happens to unmatched predictions?
You can count them as zero IoU or exclude them from distribution statistics.
Why can GIoU, DIoU, or CIoU be negative?
Those metrics include penalties beyond overlap. Poorly aligned boxes can therefore receive negative values.
How are object sizes grouped?
Records are classified using object area thresholds that you can change.
Does confidence change IoU?
No. Confidence filters predictions, while IoU measures spatial overlap.
What does the bootstrap interval show?
It estimates uncertainty around mean IoU using repeated resampling.
Can I compare models?
Yes. Include a model column and group the graph by model.
What is the Dice coefficient?
Dice is another overlap metric. For standard IoU, Dice equals 2IoU divided by 1 plus IoU.
Can I export a report?
Yes. Export CSV, JSON, chart images, or a summarized PDF report.
Does this replace an official benchmark?
No. Confirm official matching, ignored annotations, area ranges, and averaging rules before reporting benchmark results.