search

megvii-research / MOTRv2

[CVPR2023] MOTRv2: Bootstrapping End-to-End Multi-Object Tracking by Pretrained Object Detectors
Other
383 stars 47 forks source link

This can get HOTA 69.9? #2

Closed paperkaiser closed 2 years ago

paperkaiser commented 2 years ago

because i look for the training steps you provided and the use of yolox ，but can't get the HOTA of 69.9 ,So I want to ask if the code you provided can get 69.9, Thanks!!!

zyayoung commented 2 years ago

This repo is theoretically the same as the one that gets 69.9 test HOTA, except some cleanups. We do observe a relatively high instability (~2pt). Our latest run of this repo yields a 68.33 test HOTA on DanceTrack.

paperkaiser commented 2 years ago

This repo is theoretically the same as the one that gets 69.9 test HOTA, except some cleanups. We do observe a relatively high instability (~2pt). Our latest run of this repo yields a 68.33 test HOTA on DanceTrack.

thanks for your reply！

zyayoung commented 2 years ago

the use of yolox ，but can't get the HOTA of 69.9

If you add YOLOX proposal to the original MOTR repo, you may need to modify track query filtering in QIM and tune inference hyperparameters (score threshold and miss tolerance). You may refer to the "Implementation Details" section for details.

wenyuqing commented 1 year ago

This repo is theoretically the same as the one that gets 69.9 test HOTA, except some cleanups. We do observe a relatively high instability (~2pt). Our latest run of this repo yields a 68.33 test HOTA on DanceTrack.

Hi, I'm curious about what factors cause the instability as the random seed is fixed.