ByteTrack is a simple and amazing new popular SOTA method. The source code provided by the author are very insightful from research to deployment, like tensorrt, onnx, deepstream, ncpp.
One challenge I met with the official repo is that it is quite coupled with YoloX. In the meanwhile, what I need is a just simple standalone ByteTracker -- ready to use, standalone and as less dependencies as possible.
Thanks for the authors of ByteTrack, who already provides quite modularized code, I just simply further extract and made minor code (not logics) refactoring for this standalone bytetrack.
All dependencies are listed in the requirements.txt.
An example with mock videos and detectors is shown to illustrate how to use it.
From scratch, this is basically what you need
python -m pip install Cython
(this step can be ignored if you already have Cython installed)pip install -r requirements.txt
python example.py
- Make it torch independent
- Remove all the args and make the hyper-parameters explicit
- Clean up some not used imports
- Clean up the imports
- Renamed basetrack.py to base_track.py to keep names consistent
As you can see, it is more coding style and nothing to do with the logics.