search

omarabid59 / TensorflowDeepSortTracking

Tensorflow object detection with tracking based on the DeepSort algorithm
60 stars 25 forks source link

Difference from SORT #13

Open sevvalmehder opened 4 years ago

sevvalmehder commented 4 years ago

I have had chance to deeply study this project and the original deep sort project. And I review SORT and deep sort paper. Main difference between SORT and deep SORT is deep sort uses deep features of detected objects. Main deep sort repository has a features of MOT dataset as a npc file. And it loads this features. But when we do that in real time we need a deep network for feature extractor. As far as I review this project, there is no operation for feature extract. I just saw only this line for this operation: feature = [1] self.feature = np.asarray(feature, dtype=np.float32) in here: https://github.com/omarabid59/TensorflowDeepSortTracking/blob/c538fcd4a834d6b6b4c070456a79ba716161b4b7/ObjectTracking/deep_sort/detection.py#L38

Am i miss something?

JoyJulianGomes commented 4 years ago

When using the MOT dataset features are already calculated. In case of new/live videos features needs to be calculated in run time. There is not much information about this in this repository and i would suggest you to checkout https://github.com/Qidian213/deep_sort_yolov3. Specifically the the create_box_encoder from tools/generate_detections.py Hope this helps :)

sevvalmehder commented 4 years ago

After i realized this situation, i am trying to the do real time feature calculation inside this project. Thanks for referring this project. I will check.