How to actually Train with mmtracking (e.g MOT) about mmtracking

[leemengwei]

Ok， I got it. I should train with MMDetection, then use this trained model (denoted as 'detector' in MMTrack) in MMT's config file. Thanks anyway

[leemengwei]

I'm following tutorial:
Training on a single GPU python tools/train.py ${CONFIG_FILE} [optional arguments]
then I tried:
python tools/train.py configs/mot/deepsort/deepsort_faster-rcnn_fpn_4e_mot17-public-half.py
it gives me:
NotImplementedError: Please train detector and reid models first and inference with Tracktor.

Can you give me some advice? Thanks!

Also, I guess ' inference with Tracktor', the 'Tracktor' here not actually means configuration file under configs/mot/tracktor/.py right? It means any py like deepsort's sort.py and deepsort*.py. If I'm wrong, do notice me~

[leemengwei]

Ok， I got it. I should train with MMDetection, then use this trained model (denoted as 'detector' in MMTrack) in MMT's config file. Thanks anyway

Alright, here for train using mmtracking:
mmtracking/configs/det/README.md

[rvrsprdx]

Hi,

can you elaborate on how you got the training of a MOT method working?
For example for DeepSort: I don't think it is sufficient to only train a detector with MMDetection and pass it to MMTracking's config file because the Appearance Descriptor of DeepSort has to be trained as well, right?

I'd appreciate some help. Thanks.

[castrovictor]

Hi,

can you elaborate on how you got the training of a MOT method working? For example for DeepSort: I don't think it is sufficient to only train a detector with MMDetection and pass it to MMTracking's config file because the Appearance Descriptor of DeepSort has to be trained as well, right?

I'd appreciate some help. Thanks.

They use another appearance descriptor, but yes, it has to be trained. Check this and this