yihongXU / TransCenter

This is the official implementation of TransCenter (TPAMI). The code and pretrained models are now available here: https://gitlab.inria.fr/yixu/TransCenter_official.
https://team.inria.fr/robotlearn/transcenter-transformers-with-dense-queriesfor-multiple-object-tracking/
Other
108 stars 7 forks source link

The loss of the tracking decoder #20

Open xuqingyu26 opened 8 months ago

xuqingyu26 commented 8 months ago

Hello, thanks for your excellent work, and i would like to know the reason why the loss of tracking decoder is hard to decrease when applied on my custom small object dataset ?

yihongXU commented 8 months ago

Hi,

You need first check that your gt displacements are correct and might need to add data augmentation to them. Knowing that pedestrian's motion is almost constant, we designed the tracking branch to be as simple as several fc layers. For complicated motion prediction, you might need to redesign the tracking branch to have a better performance/add motion compensation.

Best,

xuqingyu26 commented 8 months ago

Thank you for your constructive feedback.