Secondary detector bboxes are sometimes larger and/or completely outside the primary bbox

I’m observing this on both DS 6.1 and 5.1 on jetson nano and dgpu.
Probably not an issue, just trying to understand how this works.

I have a primary detector, then a tracker (mot), then a secondary detector.
Most of the time the bboxes behave correctly, but whenever an object goes partially off screen
my secondary bboxes tend to go completely wrong.

e.g. (light blue bbox is supposed to be the car ‘side’)
Screenshot from 2022-11-09 19-36-36

what can I do to troubleshoot this?
Is this possibly the result of the tracker keeping a full-width image of the car internally?
Also, do I have to do my own clipping of the secondary bboxes? I thought that any secondary detection only happens within the parent bbox (+ maybe some % of margin)

Edit: same issue when the tracker is disabled.

Seems it is your model issue as you can see the same issue when disable tracker.

Ok. we do have a super custom model and issues with it are somewhat likely. I’ll check with the model guys.

I was just surprised to see that the secondary bounding box is not auto-limited to area inside of the primary bbox.
What does the secondary detector see in this case? Does it have access to the full frame?
Also: does it use the primary bbox only as a ‘hint’ of where to look for things and will ignore it in some cases?

SGIE will do inference in the bbox area. nvinfer is open source plugin. You can check if the output of the SGIE model. Please ensure SGIE works in SGIE model(process-mode).

https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinfer.html

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.