Rotated bounding boxes for nvidia deepstream: a showstopper?

Dear forum,

I have some example model that returns rotated bounding boxes in a special way, e.g. in the retinanet-format like [x,y,w,h,theta]. I’d like to integrate that into my existing deepstream pipeline that

  • A) decodes a video
  • B) inferences on the frame
  • C) draws the rotated bbox using opencv or does some other operation on the frame based on the rotated bbox
  • D) encodes the frames to a video
  • E) saves the video.

According to this forum post: (Rotated Boundingboxes) Infer live video on the jetson Platform with onnx model from the ODTK rotated bounding boxes “are not supported” by deepstream. Does that mean I will reach technical limitations that cannot be overcome? In my mind, I could write a custom plugin that accepts rotated bboxes and draws the box to the frame using opencv, right? Further, if that is not possible: Could I write a custom postprocessing plugin that calculates a minimum unrotated bbox over the rotated bbox and just draw that, so I can continue using my existing model?

Thank you in advance

Is it possible to rotate the BBox in one probe function? And then send it to nvdsosd to draw the BBox.

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