search

cheind / py-motmetrics

:bar_chart: Benchmark multiple object trackers (MOT) in Python
MIT License
1.37k stars 258 forks source link

Q: new Release #73

Closed Borda closed 4 years ago

Borda commented 4 years ago

First of all, Thanks for this very useful package. I have noticed that there is steady development on develop branch, the master is quite behind and also pip release is from mid-2018... I was wondering whether you plan a release with last changes? :]

cheind commented 4 years ago

I planned todo so before Xmas last year, but didn't find the time todo so.

On Thu, Jan 23, 2020 at 11:48 AM Jirka Borovec notifications@github.com wrote:

First of all, Thanks for this very useful package. I have noticed that there is steady development on develop branch, the master is quite behind and also pip release is from mid-2018... I was wondering whether you plan a release with last changes? :]

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/cheind/py-motmetrics/issues/73?email_source=notifications&email_token=AAAJNJJK6IU3SH2SEVVSR5LQ7FYYHA5CNFSM4KKUL7U2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4IIGSVCA, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAAJNJLDN6EK23EP4BOUHDDQ7FYYHANCNFSM4KKUL7UQ .

Borda commented 4 years ago

btw, I was running the last release on our case of 100k timestamps and it took about 370s, compared to running actual develop version which elapsed in 24s, so I am really looking forward to seeing the new release soon... :rocket:

cheind commented 4 years ago

Yes me too. I'm terribly sorry I can't act faster. I hope that I can release next week. @jvlmdr any pending things from your side? Otherwise I will try to release from develop next week. I suggest v1.2.

jvlmdr commented 4 years ago

Yep. I will try to make a short list of things to do before the release and we can discuss. Aiming for next week sounds good!

jvlmdr commented 4 years ago

@Borda Would you be OK with sending us some real data (ground-truth and predictions) for a day-long sequence with 100k+ frames? It would be useful for benchmarking.

Borda commented 4 years ago

I may share it via mail, do not want to put it public... 8-)

jvlmdr commented 4 years ago

Haha of course, no problem. If you're happy to send it to me (and I will keep it private), then use jack.valmadre@gmail.com

jvlmdr commented 4 years ago

FYI I tried out the solvers branch using the accumulator that you provided and I got some further improvement in speed (this is the cumulative time for lap.py:69(minimum_weight_matching) according to cProfile)

solver cumtime
lapmod 0.025
ortools 0.105
lap 5.166
lapsolver 6.555

However, this is still relatively small (20%) compared to the total processing time of 24 sec on your system.

jvlmdr commented 4 years ago

Wow actually.. scipy updated its implementation of linear_sum_assignment in v1.4 https://scipy.github.io/devdocs/release.1.4.0.html#scipy-optimize-improvements

branch solver cumtime (sec) line
develop scipy (1.3.3) > 30000 lap.py:6(linear_sum_assignment)
develop scipy (1.4.1) 7.158 lap.py:6(linear_sum_assignment)
solvers scipy (1.3.3) 62.333 lap.py:69(minimum_weight_matching)
solvers scipy (1.4.1) 0.120 lap.py:69(minimum_weight_matching)
cheind commented 4 years ago

an update for everyone around. If everything goes well, py-motmetrics 1.2 will be available tomorrow.

Borda commented 4 years ago

https://github.com/cheind/py-motmetrics/releases/tag/v1.2