Reproduce paper result

[trongldwork]

I clone your repo and do setting like instruction, but i cant reproduce the result in your paper, with --no_reid and full method you provide in paper, my result is

[vukasin-stanojevic]

Hello. The original results from the paper are obtained on the machine with very unusual python - conda setup. There was a lot of dependency conflicts which I somehow managed to resolve. The result was that somehow it used some packages from the base interpreter and some from the conda environment.

After submitting the paper, I tried to setup repository from the start and manage to do it with different package versions (resulting in the simple setup described in the readme file). The results obtained with this new version are slightly different (usually the difference is only on the second decimal and "unbiased" - better for some metrics and worse for the other).

Based on the numbers you present, I assume you tried it on MOT20. Below I list the results from the paper:

MOT17 validation set:	Method	HOTA	MOTA	IDF1	IDSW
BoostTrack	71.63	80.692	83.959	97
BoostTrack+	72.411	80.67	85.385	81

MOT20 validation set:	Method	HOTA	MOTA	IDF1	IDSW
BoostTrack	61.846	77.442	77.584	882
BoostTrack+	62.68	77.731	79.088	781

The results I get with different package versions:

MOT17 validation set:	Method	HOTA	MOTA	IDF1	IDSW
BoostTrack	71.532	80.7	83.875	104
BoostTrack+	72.407	80.941	85.091	83

MOT20 validation set:	Method	HOTA	MOTA	IDF1	IDSW
BoostTrack	61.742	77.464	77.454	898
BoostTrack+	62.691	77.717	79.081	787

The small differences in base results probably increase with applying the post-processing. However, the overall results are not too much different. Sorry for the inconvenience. You can consider the results you get as "normal".

Stay tuned for the new method that's build upon BoostTrack. It is implemented and tested with stable package versions :)