详细说明

[zhangmo123]

您好，看到了将FAIRMOT升级到多类别，很佩服您的工作。有关于一些细节问题想请教: 1.多类别模型的数据集是您自己的数据集，还是说将公开数据集进行组合，例如原始的FAIRMOT数据加入UA-DETRAC车辆数据 2.效果图中显示的模型的训练细节是怎样的。 如果您方便的话还是请教一下这些问题，多谢。

[CaptainEven]

@zhangmo123 训练数据很自由，可以组合公开数据集，也可以使用自己标注的数据集，只要组织成代码需要的形式即可，可以参考gen_labels_detrac_mcmot.py中的代码，方便的时候，我会放出一个小型自标注数据集的链接。

[s2244521]

@zhangmo123 , Hello, I wonder know that how the multi classes do the ReID training?

[CaptainEven]

@s2244521 Just do ReID within each class respectively will be ok.

[s2244521]

@CaptainEven , thanks, the ReID branch is like the multi class classification (ex: person, car, bicycle, ...). I wanna train multi class, if I use dataset MOT16, I need to use gen_labels_mot16_car.py to generate the labels and if the dataset is our own, use the gen_labels_detrac_mcmot.py but what is the purpose of gen_mcmot_for_detect.py ?

[CaptainEven]

@s2244521 gen_mcmot_for_detect.py is used for generating labels for detections only(parse xml and convert to txt label), and the difference is track id all set to 0.

[s2244521]

@CaptainEven , thanks for your fast reply.

[s2244521]

@CaptainEven , I wanna train the network to track person and car at the same time and I used gen_labels_mot16_car.py to generate the car label, should I need to combine MOT15(person) and MOT16(car) or change any setting in script?

[CaptainEven]

@s2244521 MOT16 is mainly for person tracking as far as i know, and it could hardly used for vehicle tracking training.

[dexter2406]

@CaptainEven , I wanna train the network to track person and car at the same time and I used gen_labels_mot16_car.py to generate the car label, should I need to combine MOT15(person) and MOT16(car) or change any setting in script?

Hi, have you figured it out how to train multi-class Re-ID networks yet?

[EdwardBY-sun]

@s2244521 @CaptainEven 请问，如果想训练自己的数据集用来实现多类多目标跟踪，数据集的制作应该参照什么格式，用什么软件或代码制作呢，如果您能回复我，我将不胜感激。