search

xinshuoweng / AB3DMOT

(IROS 2020, ECCVW 2020) Official Python Implementation for "3D Multi-Object Tracking: A Baseline and New Evaluation Metrics"
http://www.xinshuoweng.com/
Other
1.68k stars 403 forks source link

running speed on GPU #68

Closed Tsingularity closed 2 years ago

Tsingularity commented 3 years ago

Hi Xinshuo,

Thanks for the great work!

Given the amazing real-time speed of AB3DMOT on CPU, I am just curious that, what would be the speed for it on GPU?

And could you also give more details about the running time split by each module, e.g., detection, association, etc.?

Thanks!

xinshuoweng commented 2 years ago

Hi, great question! Honestly, I am not sure about the GPU speed. First of all, my code is Kalman filter based method and not a DEEP LEARNING based method, so there may be no naive way of running my code on GPU, which is why I do not know the exact speed. However, there are indeed research on developing code running kalman filter on GPU in general such as https://ieeexplore.ieee.org/abstract/document/6121397, but I did not try it yet. Would be great if someone else would like to try. :)

Also, the running time is only for the data association because detection is done off-the-shelf in my code.