Non-max Suppression - Hyperparameters


How do we tune the hyperparams in the object detection problem, IoU and Pc? Is it mainly trial and error?

Hi @Moutasem_Akkad,
Yes they are mostly hyperparameters. You have no control over the IoU directly but you can choose the threshold IoU incase that is what you mean.
There are certain tricks which can help you decrease/ increase the number of boundary boxes.
Incase you decrease the probabilty, the number of boundary boxes would increase and vise versa.
Hope this helps,

1 Like

I think Arka covers mostly. Some additions are;

In YOLO V3, P_c is defined as “objectness” to indicate how much a target object is covered by a bounding box (anchor box), and is one of outputs from the trained convolutional network.

Output features from the convolutional network are;

t_x, t_y, t_w, t_h (features[0:4]) : bounding (anchor) box shape
t_0 (features[4:5]) : objectness
t_1,t_2,...t_c (features[5:]) : class probability

IOU is used to check the coverage of bounding box. The loss function penalizes “bounding box coordinates loss”, “classification loss” and “confidence loss”. The last one is for P_c.
In the loss function, there is one parameter which can be set. That’s a threshold value that Arka mentioned. By this threshold setting, the loss function cuts off some anchor boxes whose object coverages are low, and calculates losses as miss-detection (no object) during a training phase.