Trying to solve Id switch issue of deep sort by combining feature vectors.

nwojke / deep_sort

Simple Online Realtime Tracking with a Deep Association Metric

GNU General Public License v3.0

5.3k stars 1.48k forks source link

Trying to solve Id switch issue of deep sort by combining feature vectors. #275

Open abdulw976 opened 3 years ago

abdulw976 commented 3 years ago

Hi, To improve the Id switch issue, I am trying to combine the 128 dim feature vector of deep sort with the output feature of detection layer of yolov5 and then passing the concatenated vector for cosine distance calculation but there is no change in results. Also just for experimentation when I create a dummy feature vector with all 1's still there is no change in results. Why is this happening. Please let me know of any suggestions.

Thanks

Durys commented 3 years ago

Hi - did you notice any imporovement in the ID switching? Would love to improve this feature as well.

Cheers!

ganjbakhshali commented 3 years ago

I have the same issue. in the first Idea, I change the feature extractor. Problem steel is there because of distance measuring in feature space.