search

nwojke / deep_sort

Simple Online Realtime Tracking with a Deep Association Metric
GNU General Public License v3.0
5.35k stars 1.49k forks source link

Lots of ID switches( how to fine tune the network to decrease ID switches?) #170

Open anzy0621 opened 5 years ago

anzy0621 commented 5 years ago

Hi! I'm using deep sort with a faster rcnn object detector that has been trained on vehicles dataset. I've also trained the cosine metric learning model with Veri dataset. However, when I run deep_sort the output result that I get has a lot of id-switches. Basically, the faster rcnn detections are not as stable as YOLOs detections. How do I fine tune the network? What are the major hyper parameters that I should configure to decrease the identity switches?

quocnhat commented 4 years ago

hello, have you still working on this project. in this case, I recommend you need to decrease nms_max_overlap = .75 instead of 1.0 by default. hope this help

phungthengoc commented 4 years ago

hello, have you still working on this project. in this case, I recommend you need to decrease nms_max_overlap = .75 instead of 1.0 by default. hope this help HI bro, bro co thể chỉ cho mình cách train lại Deep sort dùng cosine metric learning model không ạ ?

phungthengoc commented 4 years ago

@quocnhat HI bro, bro co thể chỉ cho mình cách train lại Deep sort dùng cosine metric learning model không ạ ?