search

nwojke / deep_sort

Simple Online Realtime Tracking with a Deep Association Metric
GNU General Public License v3.0
5.19k stars 1.46k forks source link

Input normalization (divided by 255.0) #262

Closed dangnh0611 closed 3 years ago

dangnh0611 commented 3 years ago

Hi, thanks for great work ! In the file tools/generate_detections.py, input of bgr image is feeded to the re-id encoder. The process of preprocessing only contain bgr --> rgb But in the repo Cosine Metric Learning, the preprocessing is: bgr --> rgb --> divided by 255 I saw that you has noticed that two versions contain a minor differences. So I wondering that this current implementation (without normalization) is correct or not ?

dangnh0611 commented 3 years ago

I have tried with bgr --> rgb --> divide by 255. and it shows a significant losing performance. So I think the current implementation is correct but slightly difference with the repo Cosine Metric Learning.