search

SysCV / qdtrack

Quasi-Dense Similarity Learning for Multiple Object Tracking, CVPR 2021 (Oral)
Apache License 2.0
384 stars 61 forks source link

Can I train only tracker? #126

Open YOOHYOJEONG opened 2 years ago

YOOHYOJEONG commented 2 years ago

Hi.

I trained the detector using mmcv. And, I want to use the detector checkpoint trained using mmcv for the detector of qdtrack without any additional detector learning. In this case, Can I train only tracker of qdtrack?

If I enter trained checkpoint using mmcv in init_cfg=dict(checkpoint='') in detecotr, is it the same as training only the tracker I mentioned?

Thank you.

OceanPang commented 2 years ago

Hi, you can train the model in this way, which means load a pre-trained detector but freeze the parameters and only train the tracker.

But in our experiments, joint training would give a higher performance. Anyways you can try this experiment by yourself.

YOOHYOJEONG commented 2 years ago

Thank you for your apply.

How to freeze the parametesrs? Where can I see the way? Can you check it in the qdtrack code? It's basic, but I'd appreciate it if you let me know.

Thank you.