Multi-Class MOT

[edmuthiah]

Hello,

Does TrackFormer support multi-class multi-object tracking? If so how do I go about training this?

Thanks :)

[timmeinhardt]

The MOT and CrowdHuman dataset only support a single class. But the extension to multiple classes should be very straight forward and require little implementation effort.

[JackWoo0831]

Hello, Now I want to try multi-class MOT, but I met the assert error: in util/box_ops.py: assert (boxes1[:, 2:] >= boxes1[:, :2]).all()

When I trained the single class, it happened once only because a large learning rate. but now the learning rate is proper and it doesn't work when I decrease the lr.

Have you met this issue?

Thanks.

[timmeinhardt]

I never tried to convert the repo to multiple classes. This is definitely possible and should be straightforward. But you need to change the code accordingly.

[JackWoo0831]

Hello, Now I solved this issue. To train in multi-class, you only need to modify a few codes.

1. change the 'num_classes' in models/init.py. e.g., I want to track objects of 4 classes in VisDrone dataset:

elif args. dataset==' VisDrone':# Edited
    # num_classes=1
    num_classes=4

then the classification head will change the dimension in FC layer accordingly. 2.Most importantly, when generate COCO format annotations, the category id must start with 1 and increase by one, e.g.:

annotations['categories']=[{"supercategory":"vehicle","name":"car","id":1},
{"supercategory":"vehicle","name":"van","id":2},
{"supercategory":"vehicle","name":"truck","id":3},
{"supercategory":"vehicle","name":"bus","id":4}]

[timmeinhardt]

Great! Is it working now?

[JackWoo0831]

I'm still training, and it seems good. Thank u very much for your excellent work. After some epochs I plan to add some code of showing the class name in visdom, hahaha.

[timmeinhardt]

Class name in visdom sounds good. :)

[bilalorhanlr]

Merhaba, Şimdi bu sorunu çözdüm. Çoklu sınıfta eğitim almak için sadece birkaç kodu değiştirmeniz yeterlidir.

1. Modeller/init.py içindeki 'num_classes' değerini değiştirin. örneğin, VisDrone veri kümesindeki 4 sınıfın nesnelerini izlemek istiyorum:

elif args. dataset==' VisDrone':# Edited
    # num_classes=1
    num_classes=4

daha sonra sınıflandırma kafası buna göre FC katmanındaki boyutu değiştirecektir. 2.En önemlisi, COCO biçimi ek açıklamaları oluştururken, kategori kimliği 1 ile başlamalı ve birer birer artmalıdır, örneğin:

annotations['categories']=[{"supercategory":"vehicle","name":"car","id":1},
{"supercategory":"vehicle","name":"van","id":2},
{"supercategory":"vehicle","name":"truck","id":3},
{"supercategory":"vehicle","name":"bus","id":4}]

Hello there, is it work ?