Track's box considerably different from detection's bounding box

[phungthengoc]

I trained yolov4 for vehicle include : 4 classs ( truck, car,bus,motorbike) and cosine-softmax for Deep SORT The trained-model indicate a good trainning process. But, when i have pipeline DeepSORT and Yolov4 , Track's BOX is not fit to detected-object. Its is constantly changed