Why is precision so low for yolo-nas model ?

[HissaanAli]

💡 Your Question

I trained a yolo-nas model for an object detection dataset and I noticed something weird. The recall and mAP value of the model were at 98% and 95% respectively however the precision was barely 20%.

This comes off as a surprise to me that how is the mAP and recall high, but the precision so low. Why is that ? Can somebody elaborate ?

The same model when trained with yolov5 gave precision, recall and mAP value of 94%.

Versions

No response

[ofrimasad]

Hi @HissaanAli .
It sounds like you need to tune the confidence level of the NMS.
In the post prediction callback, pass conf=0.9, for example. And see if the precision-recall tradeoff looks better