search

yihongXU / deepMOT

Official Implementation of How To Train Your Deep Multi-Object Tracker (CVPR2020)
GNU Lesser General Public License v3.0
498 stars 87 forks source link

why use DAN in inference instead of DHN? #4

Closed wangxianrui closed 5 years ago

wangxianrui commented 5 years ago

consider you have trained the SiamTPN + DHN, and why do't use the optimal assignment matrix to get the final result, but use the DAN that have not trained together?

jjldr commented 5 years ago

consider you have trained the SiamTPN + DHN, and why do't use the optimal assignment matrix to get the final result, but use the DAN that have not trained together?

Have you solved the problem? I am also confuse about DHN and DAN

yihongXU commented 5 years ago

Hi all, using DHN to do assignment during tracking is too strict. Detections are noisy and insufficient in many cases. So one-to-one assignment is too harsh for tracking, so we loose the constraint by allowing one-to-many assignment during tracking (test time). (but it should be for evaluating or training. Moreover, we use clean ground truth instead during training). DHN and DAN are two different things, yes DAN is used to do assignment in the original paper, but here we use it as an appearance extractor, just for reactivating tracks, according to appearance similarities.

Yes, we can use a better (if exist) appearance model and/or refine the appearance model on MOT dataset. I am looking forward to having your pull requests for any possible improvements.

Jacobew commented 4 years ago

@yihongXU can you please point out the details of DAN in your paper? I can't find it anywhere. I am a little bit surprised that you do not use fixed DHN in inference like what DeepMOT does in training.