A question about object association at different time steps

[lebron-2016]

Dear author,

In testDis.py , I found that you used Kalman filter when predicting the state of the target in the next frame, instead of the tracking_dis predicted by the model. Why can't we directly use the offset predicted by the model? What is the significance of the offset(or inter-frame displacement) of the target here?

Thanks!!

[DonDominic]

Thanks for your attention to our work.

In our previous experiments, we found that incorporating a Kalman filter results in more stable tracking performance compared to using only inter-frame displacement. Consequently, we released this version that combines both the Kalman filter and inter-frame displacement. Additionally, we provided a test using the Kalman filter (SORT) alone in the 'testTrackingSort.py' script. You might find it helpful to run this testing script to see the impact of inter-frame displacement clearly.

[lebron-2016]

Thanks for your attention to our work.

In our previous experiments, we found that incorporating a Kalman filter results in more stable tracking performance compared to using only inter-frame displacement. Consequently, we released this version that combines both the Kalman filter and inter-frame displacement. Additionally, we provided a test using the Kalman filter (SORT) alone in the 'testTrackingSort.py' script. You might find it helpful to run this testing script to see the impact of inter-frame displacement clearly.

Thanks for your quick reply!

I have three more questions that I would like to confirm.

1. Are you just using the inter-frame offset as a result correction? (as shown below)

2. Are you only using Kalman filter when predicting the state of the object in the next frame?

3. About testTrackingSort.py, is it equivalent to modifying testDis.py as follows？

Thanks!!

[DonDominic]

Yes, for questions 1 and 2, in this setting we use inter-frame displacement as result correction and incorporate the Kalman filter to conduct state updates. For question 3, that's correct.

[lebron-2016]

Yes, for questions 1 and 2, in this setting we use inter-frame displacement as result correction and incorporate the Kalman filter to conduct state updates. For question 3, that's correct.

Got it! Thanks!