( Jetson tx2, DeepStream 6, YOLOv4 tiny)
At the deepstream app examples (such as test-1…4) in order to use the model inference (the NN) output,
as I understood, we should add a probe to the nvinfer (pgie) plugin, and extract the detection from the relevant metadata variable.
And if we would like to draw the bbox on the frame we need to attach another probe (a callback function) to the
source of the ‘nvdsosd’ plugin (osd), which has access to metadata detection’s, and the frame on which he should draw the bbox.
First question: Why I cant draw the bbox already at the callback function of the nvinfer (pgie) plugin?
After long research I realized that YOLOv4 tiny needs another configuration at the nvinfer’s plugin file
which is the “parse-bbox-func-name” param.
Second question: Why do we need this configuration, since in the examples provided (test1…4 and others)
the ‘nvdsosd’ plugin gets the detection bbox directly from the metadata (obj_meta.rect_params) without extra parsing, what Im missing here?