franktpmvu
commented
1 year ago It depends on how often you experience occlusion, if all of video without occlude, inference speed will equal to original method.
Based on my experience, VOT dataset will decrease 33% inference speed (means 0.66 x original fps). each neighbor need reverse track n frame.
For now, each neighbor use only one for loop to get reverse tracklet. If your SOT method support multiple kernel, you can speed up this for loop: https://github.com/franktpmvu/NeighborTrack/blob/89aa0781c5b59ac570e3c1c47cca5b1dd6a5f945/neighbortrack.py#L150 it will be helpful.
about hardware-friendly, because our method depends on your SOT method, if your SOT method is a hardware-friendly method like NCC tracker, then our method are hardware-friendly method.
update
In my new test on paper, the table of fps / tau show in below
Does post-processing take a long time and is it hardware-friendly?