phil-bergmann / tracking_wo_bnw

Implementation of "Tracking without bells and whistles” and the multi-object tracking "Tracktor"
https://arxiv.org/abs/1903.05625
GNU General Public License v3.0
818 stars 196 forks source link

how is the speed of this tracker? It seems MOT challenge site gives 0.6 Hz which is rather low #11

Closed rambleramble closed 5 years ago

rambleramble commented 5 years ago

Thanks for this great work,

According to the MOT challenge website:

https://motchallenge.net/results/CVPR_2019_Tracking_Challenge/

The speed is only 0.6 Hz, is it a true reflection of this repo?

Do you have a more recent analysis on the speed performance?

Thanks in advance!

timmeinhardt commented 5 years ago

I just tested Tracktor++ on the CVPR 2019 test set and it took 1667 seconds to finish all 4 sequences. I updated the MOTChallenge webpage accordingly.

rambleramble commented 5 years ago

thanks for the update!

My primary interest is the speed of the performance.

If possible, could you provide some insights on how to improve the speed? Or which module (such as re-id) that can be further optimized in order to gain a further speed-up?

Also, if the number of detected object (peoples) is reduced (say only detect major characters in an image by setting a higher detector threshold ), will this improve the process speed?

timmeinhardt commented 5 years ago

We did not evaluate the runtime of the different steps in our Tracktor algorithm. However, there are multiple things one could do to speed it up. Our object detector is called multiple times. I think the number of calls per frame could be reduced. Running an object detector with a different backbone, e.g., ResNet34 instead of ResNet101, is a possibility if top tracking performance is not essential.

Changing the trajectory killing or bounding box initialisation thresholds should change the number of detected objects but I think the potential for reducing the runtime is comparatively small.

But again, this is something that has to be evaluated to be really sure.

timmeinhardt commented 5 years ago

Can we close this issue?