Too fast to ID switch

[zhouzw87]

Hi, I think deep_sort is a great job. If my detection-module not detected object in one frame, but next frame detected object, ID will change the new ID. The occlusion also is the same situation, one object was occluded in a frame, When it appear again, it's ID will change.

[ninghongbo123]

yes，I have the same problem as you. And I guess that it will work better if modifying some hyper parameters ,but firstly it is necessary to have a comprehensive understanding of this project. And I am dong so.

[Akhtar303nu]

@zhouzw87 @ninghongbo123 any update about hyper parameter to reduce id Switching Thanks

[ss199302]

@ninghongbo123 hello！Do you know how to modify the parameters?

[ninghongbo123]

I had modified these parameter or operating logic : 1.threshold-replace 0.2 by 0.25 2.average the top five features of the track as final distance OR average the top 3rd to 5th features, i.e. remove the top 2 features 3.Every frame will carry out these steps that include tracker.predict() and tracker.update()
regardless of whether or not having a detection. 4.For matching cascade, trackers whose age<=5 are gathered together to associate at the same time @ss199302 @Akhtar303nu

[ss199302]

@ninghongbo123 Thank you for your answer. How do you modify point 2 in your code?

[ninghongbo123]

In  nn_matching.py file,  modifying  corresponding code as  below:
      distances.sort(axis=0)    # sorted firstly
      return distances[2:5].mean(axis=0)  #average  2-5  features

@ss199302

[ss199302]

@ninghongbo123 Thank you for your reply. I have made some modifications according to your code, but the effect is not very good. Is the weight model you use well trained by the authorities?

[ss199302]

@ninghongbo123 Hello, I'd like to try the fourth point you mentioned. Is it to modify the max age? If not, how to modify it? Why is that? I don't understand. Thank you

[ninghongbo123]

Any modification you made depends on your application. You need to evaluate it after you have modified some parameters .
I think these two parameters are irrelevant . The max age is just your expectation that the tracker should be removed . And my fourth point aims to give trackers whose age<=5 equal possibility to obtain a detection. @ss199302

[ss199302]

@ninghongbo123 How do you modify the code for the fourth point? Can you tell me? thank you very much

[ciwei123]

@ninghongbo123 Hi, After you modified the parameters, what data set did you test? What is the effect? And the mars-small128.pb provided by author is trained just through the Mars datasets?Thank you very much.

[Mps24-7uk]

@ninghongbo123 how could I modify point 4 (age<=5 ) in the code.

[anzy0621]

@ninghongbo123 how to modify the point 3 and 4 that you mentioned above?

Thank you in advance!

[hongsamvo]

@ninghongbo123 thank you for your suggestion. I am working with vehicle counting project. My detector is not good, some frame it lost the detection and the ID of tracking is not switched. but it lost the tracking bbox after lost detection bbox. Do you have any solution to keep the tracking bbox after lost detection bbox, it will make bbox is smooth and the counting result will better. Thank you very much

[ninghongbo123]

@hongsamvo
hello !
the performance of detector has a important impact on final result of tracking , so it's better to improve the capabilityof detector as soon as possible . I don't know what your specific application is , but as far as i know , vehicle is far more easier to be detected and be tracking than pedestrian . If it lost detection bbox , you should make sure that the information of lost objects can be preserved until next step . There are some methods to try it , including updating the result of KalmanFilting on the assumption that there is no any lost frames , or adopting a robust feature extracter and raise the weights of feature score , and so on .

[hongsamvo]

@ninghongbo123 Thank you very much for your suggestion. I found a solution that similar with your advise that they assume that there is no lost frames and keep tracking bbox until the id is deleted(https://github.com/abewley/sort/issues/55). I am working counting vehicle. My method base on the intersection between a line and the line which connect the center bbox of the same id in the 2 consecutive frames (https://github.com/guillelopez/python-traffic-counter-with-yolo-and-sort). This solution helped me to improve the counting result. The performance of my detector is not really good and I am trying to improve it.

[phungthengoc]

@hongsamvo HI ban, ban da train lai duoc deepsort cho Vehicle chua a. ? minh cung dang lam project tuong tu nhung train lai cho vehicle bang cosine metric learning thi minh gap nhieu error qua.

[hongsamvo]

@phungthengoc Hi bạn, mình cũng dùng repo cosine metric learning để train, không rõ là bạn gặp error gì?

[kurkurzz]

I believe point number 4 can be modified by referring to this comment:

https://github.com/nwojke/deep_sort/issues/256#issuecomment-902463018