Custom object detector depth and object dimensions

So I’m using a custom object detector (darknet yolov4tiny optimized with tkDNN) to detect fruits in the wild (i.e. in their natural environment).
I’m accessing per object spatial information using the yielded object data.

Question is - how does the given distance, and therefore object dimensions, are being computed? does it referenced to the surface distance (i.e. the outer part of the object closest to the camera) or does it referenced to the centroid distance of the real 3d object.

I’m getting a bias which indicate that the dimensions of a given object is being calculated with respect to the closest point (to the camera) of the object. If that is true, is it possible to set the distance estimation to be that of the real 3d centroid?

System information:

  • ZED sdk 3.7.2.
  • CUDA 11.5.
  • Ubuntu 20.04.
  • ZED2i.
  • Python 3.7

Hi yotamra,

we use the centroid of visible surface points to compute the 3d position of each object for now. You could fix the bias if you know the exact dimension of the detected object.

1 Like