TensorFlow-2.x-YOLOv3 icon indicating copy to clipboard operation
TensorFlow-2.x-YOLOv3 copied to clipboard

Parameter Tuning

Open AMYEY opened this issue 4 years ago • 3 comments

Thank you for your tutorial, it has helped me. However if any object is occluded in a frame and after few frames the occluded object reappears, the object ID changes. This is happening quite a lot. ID switches occur every time two persons clash. Any suggestions to fine tune the parameters? I would also like to understand the max_iou_distance and n_init parameters.

AMYEY avatar Jul 06 '21 07:07 AMYEY

This is the fault of the tracker, I spent around 3 weeks understanding how the tracker works, and so on... If you are ready to spend 3 weeks just to understand the tracker, go for it. But I will tell you if you want to make it work better, you need to train your own tracking model or rewrite tracker code that tracker would be able to revive dead tracks and compare new tracks against dead tracks by using model predicted features.

pythonlessons avatar Jul 06 '21 07:07 pythonlessons

Thank you for fast response. I need to understand the tracker parameters, any suggestions from where I could understand these...for example the max_iou_distance. Or any resource you could suggest in understanding of how tracker works would be really helpful. I cannot find much online.

AMYEY avatar Jul 06 '21 07:07 AMYEY

I can't give you one sentence answer, if you need understanding you need to read the code. max_iou_distance is used to calculate distance between two objects in Kalman filter

pythonlessons avatar Jul 08 '21 12:07 pythonlessons