search

ifzhang / ByteTrack

[ECCV 2022] ByteTrack: Multi-Object Tracking by Associating Every Detection Box
MIT License
4.5k stars 858 forks source link

Backbone Change #75

Open abhigoku10 opened 2 years ago

abhigoku10 commented 2 years ago

@iFighting @ifzhang Hi thanks for sharing the code base and for great work , i have following queries

  1. Can we use ur training pipline to train other classes like bags,cars and use it for byte tracking
  2. Currently the model is trained using Yolox-X backbone can we change to Yolox-L/s/m backbone ? if so what changes have to be made in the code base
  3. Can Byte Track be used for multiclass tracking of objects like cars , bags Thanks in advance
ifzhang commented 2 years ago
  1. Yes, you can add more classes in the detector.
  2. You can use the exp files in exps/example/mot to change to other backbones.
  3. Yes, first you need to train a multi-class detector and run tracking for each class.
abhigoku10 commented 2 years ago

@ifzhang have you trained the detector on different backbone ? can you provide the pretrained weigths of the model if available on google drive or one drive ??

zhangby2085 commented 2 years ago

@ifzhang, may I ask, for the detector, in which file and location to add more classes in the detector? would like to use the original yolox_s weight file. Thank you very much.

wenchima commented 2 years ago
  1. Yes, you can add more classes in the detector.
  2. You can use the exp files in exps/example/mot to change to other backbones.
  3. Yes, first you need to train a multi-class detector and run tracking for each class.

@ifzhang This is good news! Thank you! Just curious about the third item, you mentioned "run tracking for each class". Does that mean the model could only support single-class tracking in a single tracking inference or we can do multi-class multi-object tracking inference at the same time?

OceanStarHit commented 2 years ago
  1. Yes, you can add more classes in the detector.
  2. You can use the exp files in exps/example/mot to change to other backbones.
  3. Yes, first you need to train a multi-class detector and run tracking for each class.

Thanks for your kind answer for us, developers. I desired and expected those, but was not sure. Now I have confirmed.. Thanks.

OceanStarHit commented 2 years ago

Hi, ifzhang. Could you guide and share to me some more if you have time at any time? I'd like to discuss more with you. Thanks a lot.