how to set iou accroding val result?

[jaffe-fly]

after val get

ap_class_index: array([55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 86])
box: ultralytics.utils.metrics.Metric object
confusion_matrix: <ultralytics.utils.metrics.ConfusionMatrix object at 0x72f7f1eedc90>
curves: []
curves_results: []
fitness: 0.9704991125053236
keys: ['metrics/precision(B)', 'metrics/recall(B)', 'metrics/mAP50(B)', 'metrics/mAP50-95(B)']
maps: array([    0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,
           0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,
           0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819,      0.9836,     0.99189,     0.98066,     0.98594,     0.96873,     0.98867,     0.97634,     0.95952,     0.98118,     0.95849,     0.96499,     0.93139,      0.8764,     0.91598,     0.90748,     0.89591,     0.92786,
           0.99026,       0.995,     0.99424,     0.95875,     0.99347,     0.99328,     0.99075,     0.97703,     0.99108,     0.98247,     0.99384,     0.98745,     0.97773,     0.96819,     0.99355,     0.96819,     0.96819,     0.96819,     0.96819,     0.96819])

why maps has so many values?
in mAP50-95 , which iou value will get hightest map?

[UltralyticsAssistant]

👋 Hello @jaffe-fly, thank you for your interest in Ultralytics 🚀!

If this is a 🐛 Bug Report, could you please provide a minimum reproducible example (MRE) to help us debug the issue? This will ensure we can provide the most accurate assistance possible.

If this is a ❓ Question about custom training or validation results, providing additional details like your training configuration, dataset examples, and logs would be extremely helpful. It’s also a good idea to verify that your training process follows best practices for model training.

Lastly, remember that an Ultralytics engineer will review your discussion soon to provide more specific assistance. Stay tuned! 😊

[glenn-jocher]

@jaffe-fly to adjust IoU based on validation results, you can set the iou_thres parameter during validation. For example: model.val(iou_thres=0.7). The maps array contains mAP scores for all classes (including potentially unused ones), where higher mAP typically occurs at lower IoU thresholds like 0.5. For detailed insights, review the PR_curve.png generated in your validation output directory, which shows precision-recall tradeoffs across IoU thresholds.

[jaffe-fly]

so this mean iou=0.5 ,this model is best?

[glenn-jocher]

@jaffe-fly the maps array contains mAP50-95 scores for each class (one value per class). While mAP50 shows performance at IoU=0.5, the highest mAP typically occurs at lower IoU thresholds like 0.5 since higher IoUs (e.g., 0.95) require stricter box matching. To find the optimal IoU for your use case, consider your application's precision needs - use lower IoU (0.5) for general detection or higher IoU (0.7-0.95) for precise localization. You can set IoU during validation via iou=0.x in model.val() or CLI. For details, see our metrics guide: https://docs.ultralytics.com/guides/yolo-performance-metrics/

[Y-T-G]

That's mAP for each class.

The iou that you pass during val or predict is for NMS. It's not the same thing as the IoU shown here.