trackerType SHORT_TERM_KCF

raviks

Hi,
We recently purchased 2 OAK-D devices and trying out the ObjectTracker node functionality using tracker type as SHORT_TERM_KCF which is based on skipping few frames from running the detection node for the best accuracy. But there was no example/reference code available anywhere in your sdk documentation nor in the gtihub depth-ai-experiments.

It will be a great help if anyone can throw some light on how to utilise tracker type SHORT_TERM_KCF.

Thanks,
Ravi.K

erik

Hello raviks , I will add it to my TODO and probably finish it this week.

raviks

Thanks a bunch Erik. Will look forward to your example....

raviks

@erik any updates on this please. Thanks

erik

Hello raviks , after some evaluation it looks like KCF is too slow to be very useful. What's the model you would like to use it with?
Thanks, Erik

raviks

@erik since SHORT_TERM tracker type helps to reduce running face detection for every frame, we are looking for any example using SHORT_TERM tracker type tracklets with either SHORT_TERM_IMAGELESS or SHORT_TERM_KCF is fine.

We have multiple models that we need to run in sequence for our use-case based on accurate people tracking and hence achieving a overall FPS of around 10 has become very challenging for us. i am assuming that using SHORT_TERM tracker type will help us achieve good accuracy with lower FPS.

Thank You,
Ravi.K