Train reid and detector for custom objects /labels ( not people/cars)

[whiteplatin]

Hi,

is there a guide on how to set up a custom dataset to train a reid model, including which annotation format to label the data as (e.g. Pacal, COCOjson etc.) ? Even a general tutorial would be fine, looking on the net , I was unable to find anything except for MOT and other similiar datasets.

The goal is to use this tracking algorithm for custom objects.

Thanks

[timmeinhardt]

We do not provide a detailed tutorial. But all the information can be found in the repo, e.g., for reid look at the reid_mot_to_coco_gt.py file.

[whiteplatin]

Hey thanks, can you give more details about the names of the annotation fomrat ?

Reid mot to coco ?

I mean I already have my dataset in coco format (but without the frame_id, which I guess is essential for reid training), Ive never heard of mot annotation, would this be the way to create it ,https://motchallenge.net/instructions/?

[timmeinhardt]

If its already in COCO you can look at reid_coco_to_coco_gt.py. But generallt I dont know how your reid files are formatted. But if you look at the files and understand how it is supposed to look like you can write yourself a similar convertion script if necessary.

[whiteplatin]

I do not have reid files nor do I know what they are , if you have a link that explains what it is, that would be great.

reid_coco_to_coco_gt.pydoes not seem to exist:

[timmeinhardt]

Sorry, the reid_coco_to_coco_gt.py file is not yet published. But you can find everything in the reid_mot_to_coco_gt.py file. It shows what needs to be done to train reid. You can also look at this https://pytorch.org/vision/stable/index.html. Its the package we use for training reid.

[whiteplatin]

Thanks very much for your time and help. I will have a read through it.