Hi Wojke:
Thank you for your Min-Cost Flow python code in github. I tried the same algorithm on matlab, but, I really want to discuss with you a problem in Min-Cost Network Flow tracking.
Generally, there exists several False Negatives/Occlusions in the video. So it is imperative to set transition links between longer temporal gaps, e.g. 5 frames.
But it also dampens tracking performance, when I did so, like in TUD-Crossing sequence with ACF detector, the tracker usually has big delayed initializations for some tracks and so many false negatives, for example, it only have 2-3 tracks in the first 10 frames, although there are 7 targets in first 10 frames, even I set the entry cost to be a smaller value near zero, the problem exists. Assuming I already have a very good pairwise cost function trained offline.
But, tuning the entry/exit cost parameters is desperate. Do you have any suggestions on how to handle this problem?
Thank you for your time and help!
Hi Wojke: Thank you for your Min-Cost Flow python code in github. I tried the same algorithm on matlab, but, I really want to discuss with you a problem in Min-Cost Network Flow tracking. Generally, there exists several False Negatives/Occlusions in the video. So it is imperative to set transition links between longer temporal gaps, e.g. 5 frames. But it also dampens tracking performance, when I did so, like in TUD-Crossing sequence with ACF detector, the tracker usually has big delayed initializations for some tracks and so many false negatives, for example, it only have 2-3 tracks in the first 10 frames, although there are 7 targets in first 10 frames, even I set the entry cost to be a smaller value near zero, the problem exists. Assuming I already have a very good pairwise cost function trained offline. But, tuning the entry/exit cost parameters is desperate. Do you have any suggestions on how to handle this problem? Thank you for your time and help!