Is it possible for TensorRT model from here? about yolo_tracking HOT 4 CLOSED

Thank you!

I am not familiar with TensorRT but from what I understand you should be able to serialize the Deep Sort weights in the same way you did with the Yolov5 weights, of course taking into consideration the architecture differences. The network for deep sort is just comprised of convolutional layers followed by residual blocks and a dense layer so it should be easier to serialize than Yolov5.

Then you could deserialize them in a c++ ported version of track.py for real-time tracking

from yolo_tracking.

@mikel-brostrom Thanks. I have trained yolov5 for my custom object detection, Similarly do i need to train 'tracking model' deepsort for tracking my custom object?

from yolo_tracking.

Yes, you need to train the deep association metric. Ideally on a large-scale re-identification dataset containing the classes you want to track

from yolo_tracking.

我试了yolov5_deepsort的tensorrt版本，大概15ms一帧

from yolo_tracking.