Please provide complete information as applicable to your setup.
• Hardware Platform (Jetson / GPU) - GPU • DeepStream Version - 6.3 • JetPack Version (valid for Jetson only) • TensorRT Version • NVIDIA GPU Driver Version (valid for GPU only) - 550.107 • Issue Type - question
I’ve the pipeline -
camera → PGIE (yolo) → SGIE (face detector) → tracker → osd
Currently, the tracker is tracking all the detections - 80 classes of yolo + face bboxes. So it gives tracking ids to all classes of yolo (chair, etc) and faces.
Instead, I want the tracker to track detections coming from SGIE only. Or track based on class-ids, eg. track only person class from yolo and all detections from SGIE. How to do that?
I saw the tracker config tensor-meta-gie-id, but it is only applicable when input-tensor-meta is added by a preprocess plugin. Does it mean I’ll have to do something like add a preprocess plugin to assign a gie id at tensor meta level? Can’t the deepstream tracker track based on specifc nvinfer-gie-ids??
Using pre-cluster-threshold=1.5 for all classes except class 0 (person) means that those classes will be filtered out, since confidence threshold for any class cannot be greater than 1.
Forgot to mention this, while these class attributes work when network-type=0 (Detector), I’ve set all the GIE network-type=100, because I need a custom model output parser. So when I set the class attributes in GIE config file, there’s no impact in the tracker output.