Mike

The more objects to track the longer the time to generate the appearance embeddings and the association process to finish.

If you are interested in a single object you can discard the rest of the tracked objects, lowering the computational load. This would require some custom modifications.

You could use the bounding box prediction generated by the tracker for the specific object of interest, compare to it the rest for example by using IoU and finally discard...

Could be more suitable. The KF in OCSORT is designed for non-linear moving objects.

> > The more objects to track the longer the time to generate the appearance embeddings and the association process to finish. > > Yes, I can discard the rest...

This shouldn't be an issue when installing as specified under README 😄. Notice you have to install the custom Ultralytics package that this repo is pointing to

Yup, I see. I will discontinue the support for python3.8

Done @flonix8 !