Regarding details of MOT

[Sudarshan Kamath]

Hi,

Hey guys!

I was reading the paper on BDD100k and I saw this under the MoT section:

"For the

detection and MOT models, we use a modified version of

Faster R-CNN [27] with a shared DLA-34 [36] backbone.

The implementation details of the tracking model are shown

in the supplementary materials."

I couldn't find any links to the supplementary materials. I am especially interested in how the re-ID task was performed. 

Are you using a similar approach to fairMOT where you use kalmann on re-ID features coming as an output from the FRCNN? If so how is the kalmann updated?

Awaiting your response

Thanks

[Fisher Yu]

Hi,

You can find our latest code for tracking on BDD100K here: https://github.com/SysCV/qdtrack. It doesn't use kalman filter for re-ID. 

BTW, we moved to https://github.com/bdd100k/bdd100k/discussions for discussion.

Cheers,

Fisher