DeepSORT features

Question

I'm reading about Detection and Tracking algorithms and I'm unclear about the DeepSORT algorithm:

1. How does the DeepSORT algorithm gets the features? Does it "hijack" the feature vector from the upstream detection algorithm? (such as YOLO? or others?). This seems unreasonable to me, since not all methods would make it easy to get the feature vector.
2. Does it create features on its own, using a pretrained CNN network? This seems to make sense, and also makes it independent from the Detection algorithm. So I would imagine that DeepSORT gets the bounding-box for each object, and then would need to do some image-preprocessing on its own? (such as crop/resize the part of the image that's related to the BB?)

Thanks.

Answer 1

The answer is (2). This makes sense because the features created for the Detection task are different from the ones needed for the Tracking task. It is one thing to detect a car, and there might be multiple cars. You might not need features such as make, model, and the specific numbers on the license plate to determine "yes - this IS a car". But to differentiate one car from another (even passing by each other) such features might be needed.

This is not generally true (there are models that do detection/tracking/re-id) using a single features-network, but DeepSORT isn't one of those.