Qidian213

Now i am focus on tracking , some day may improve it, deep sort ‘s features extract ability not very good . wish more commication.

Many papers of Person Re-Identification can be reference to , for training method 'Triplet Loss with hard negative mining ' can be better than deep sort use ' classification '.

I afraid that `it can track cars and person at the same time,but trafficlight 、bicyclist are also tracked!!!! ` because of the model you trained was not performs well. And...

https://github.com/Qidian213/deep_sort_yolov3/issues/7

First ，there has a bug i forgot to update on github . it should be `if not track.is_confirmed() or track.time_since_update > 1:` in demo.py file . Second for the performance...

I think a better loss function is more effective than increase the size of feature vector .For different people have same apperance , same person have varying apperance . In...

For using indoor ,a dataset bese on indoor case may better .

why not have a try ?

Sorry , I dont like to write it , so no report or blog.

checkpoint file can be found in [https://drive.google.com/drive/folders/1m2ebLHB2JThZC8vWGDYEKGsevLssSkjo](url),and I have not tried tensorRT.