Yolov5 NMS calculation

[krishnakanagal]

I want to understand why the object confidence score is multiplied with the class confidence score in the non_max_suppression in yolov5?

x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf

Is there any benefit in doing this? Why is the class confidence score not used as is?

[glenn-jocher]

@krishnakanagal class confidence times object confidence is the total confidence that is used for NMS.

[nrupatunga]

@krishnakanagal It's the common practice and with the logic that if any one of the confidences is low, we make sure to deprioritize them during NMS and hence multiplication obj_conf * cls_conf

[TrueWodzu]

What do you mean by "deprioritize them", they are simply removed. This is counter-intuitive because if we are passing conf_thres as a parameter, then we are expecting to prune all boxes that are below that threshold, not some other threshold that is being calculated under the hood.