It seems so common to use "dumb" trackers right behind object detection models. Why? Is it not preferable to try to model the actual physical process this early in the pipeline?

For instance in the case of tracking fixed objects from moving vehicles. We can mesure the speed and direction of the vehicle from other sensors fairly accurately and use a better model of the movement. Why not put that and the noisy bounding box out of the model *directly* in a filter?

#machinelearning #algorithms