phil-bergmann / tracking_wo_bnw

Implementation of "Tracking without bells and whistles” and the multi-object tracking "Tracktor"

Home Page:https://arxiv.org/abs/1903.05625

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

frame_split and overall metrics

amoskalev opened this issue · comments

Hello! How the metrics are calculated when the frame_split is not [0, 1]?

For example, if I set frame_split=[0, 0.25] then the tracking happens only on the first quarter of each sequence, right? How the overall metrics are then computed? What tracks are used in the sequence for the remaining frame_split=[0.25,1]?

Thanks!

In that case the MOT metrics are only calculated for this part of the sequence. The ground truth of the rest is not considered at all.

@timmeinhardt Hmm. When I run tracktor on full MOT20_train the metrics are all ok, but if I run it with frame_split=[0.3, 0.6] then IDF is only around 3% and MOTA even becomes negative.. do you have any ideas why something like this happens?

I am not sure why this is the case. You could use the write_images argument to output qualitative results and see if the tracking or its evaluation fails.

I plotted the sequences and I think the problem may be in how results are recorded.

Frame split works just fine when it's of the form [0, x]. However, if you specify the beginning of the split as non-zero, then the results are misaligned.

Yes, looking at the code this makes total sense. I pushed some changes that should fix the issue.