Problems running tracker with DeepSORT

[natashasoon]

Hi, thank you for the great work on this tracker!

I am trying to run example 5 with DeepSORT, but have some issues relating to OpenVINO. I am using OpenCV 4.5.4, and the tracker is able to run properly when opencv is built without the inference engine. However, when I built opencv with inference engine support (openvino 2021.4.1), I am not able to run the tracker even with useDeepSORT = false. When set to false, the yolo weights are loaded, but i run into a Segmentation fault (core dumped) error. On the other hand, when useDeepSORT = true, I get this error instead: ./MultitargetTracker: symbol lookup error: libmtracking.so: undefined symbol: _ZN2cv3dnn14dnn4_v202110043NetC1Ev.

I noticed from your response in issue #276 that you used opencv built with OpenVINO 2021.2.185, which i understand is compatible with OpenCV 4.5.1. Could the issue be with my opencv version?

[Nuzhny007]

Hi!
Thnx for describing this bug. I'll try to use some modern re-id architecture without OpenVINO

[natashasoon]

Thank you so much!

[Nuzhny007]

Sorry for the long answer. I have added support one model, you can see changes here: https://github.com/Smorodov/Multitarget-tracker/pull/430/files

[natashasoon]

Hi, thank you so much for taking the time to make the changes. It works well for me. However, I noticed that your calculation for the cosine distance is now 1 - sim (where sim is the cosine similarity), instead of the original 0.5 * (1-sim). May I know the reasoning behind this? The original calculation seems more suitable to me since it scales the distance to a range of [0,1], while the new calculation has a range of [0,2].

[Nuzhny007]

Hi!
Thnx for remark!
I founded it here: https://github.com/bitzy/DeepSort/blob/db8b64e594c97493a0c289a7ce22fa69530b9075/src/matching/nn_matching.cpp#L158
And in some another places. Do you think that it wrong?

[natashasoon]

Sorry for the late response! The calculation is not necessarily wrong, just that it seems to make more sense to scale the cosine distance metric to have a range of [0,1] since most (if not all) of the other metrics like the jaccard distance have the same range. Since the distance can be a weighted combination of all the metrics where the weights sum up to 1 (judging by examples.h), it makes the most sense to me if all the metrics have the same range of [0,1].