esdolo / immortaltracker Goto Github PK

[2111.13672] Immortal Tracker: Tracklet Never Dies https://arxiv.org/abs/2111.13672

Hi, I'm very interested in your work ImmortalTracker, at the same time I'm following tusen's SimpleTrack, I wonder what is the difference between these two codebases? Except that the life cycle of the tracker in ImmortalTracker is infinite, I didn't find much difference elsewhere.

Hi @ImmortalTracker,

Many thanks for your work!

I would like to reproduce you work and was wondering if you instructions for re-training?

I assume you model needs training and not only the base detector. Please correct me if I am wrong here.

Best wishes

Hi, thank you very much for sharing. However, when validating the nuscenes data for 20hz, the following error occurs:
Traceback (most recent call last):
_File "/home/wsd/anaconda3/envs/simpletrack/tib/python3.7/site-packages/nuscenes/eval/tracking/evaluate.py", line 270，in render classes=render classes)
File "/home/wsd/anaconda3/envs/simpletrack/lib/python3.7/site-packages/nuscenes/eval/tracking/evaluate.py", line 84,in __init_verbose=verbose)File "/home/usd/anaconda3/envs/simpletrack/lib/python3.7/site-packages/nuscenes/eval/common/loaderspy", line 48,in load_prediction"Error: Only <= %d boxes per sample allowed!" % max_boxes_per_sample
AssertionError: Error: Only <= 500 boxes per sample allowed!_
And I see that the data in your paper seems to be 2hz too, how should I calculate it?

Hi, I'm interested in your work, but I have some questions I want to discuss with you.

As oc-sort said, the noise of kalman filter will increase with t^2 as long as there is no observations in update step, which makes the prediction of kalman filter not reliable anymore. So in my experiments in MOT17 dataset, simply increase MaxAge of lost trajectories will cause more wrong matches.

What do you think of this question? Looking forward to your reply!