The confidence value from object meta of detected object is returning weird values eg. -0.10000000149011612


I am using Deepstream SDK 5.0 and have implemented GSTPipeline in which I am using PeopleNet as my primary detector for person and face detection. I have specified pre_cluster_threshold as 0.25 for the detection model.

Now, I am trying to filter out detected faces and persons bbox based on their confidence value. But when I am extracting the bbox confidence value from its object_meta(NvDsObjectMeta object) it return almost same value for all bboxes and values are also weird like -0.10000000149011612.

While the examples mentioned on the forums compares confidence value to some float values like 0.4, 0.5 …etc which is nothing but the decimal representation of confidence percentage. I am referring deepstream test app 4 for this.

Can anyone suggest if my understanding is correct and if I am using the right approach. If more information need on for further clarification let me know



I am facing same issue.

Basically, we would like to fetch the precision with which the object is detected / classified. Is confidence value from object_meta (NvDsObjectMeta) the right data structure to fetch the same? In our case, we have multiple scenarios in gstpipeline, where we need to filter as per higher threshold and some places need lower threshold. So we have set lower threshold value in the config file and trying to check precision of object detection/classification using confidence value, where a deviation is needed.

Any help is appreciated. Thanks.