Recently I read an interesting paper "Real-time Compressive Tracking" which applied compressive sensing on object tracking to dramatically reduce the calculations and showed remarkable results. Kaihua Zhang (the author) kindly had released the source code(both Matlab and C++ versions) of the algorithm on his webpage.
A python binding
Personally I prefer to test algorithms with python, hence, I have tried to build a python binding of compressive tracking and tested it on my Mac.
Some results are as below:
-
Kitesurf
-
Bolt
and the source code can be found from this link.
More details
This section will describe the main steps how I wrapped the C++ version of compressive tracking as the python binding and is not important in the aspect of usage. In case you are interested in the procedures, please keep reading.
The class CompressiveTracker in compressive_tracker.cpp provides two public methods init(Mat& _frame, Rect& _objectBox) and processFrame(Mat& _frame, Rect& _objectBox),
where _frame and _objectBox are the video frame and object box, respectively.
I added two public methods init_wrap(vector<vector<uint8> > &_frame, vector<int> &_object_box) and process_frame_wrap(vector<vector<uint8> > &_frame, vector<int> &_object_box) to wrap the methods init and processFrame, which expect vector array arguments instead of Mat or Rect ones.
Next, the cython files wrap.pxd and wrap.pyx were created to wrap the C++ class CompressiveTracker into the Python class CyCompressiveTracker.
Besides, setup.py was added for compilation.
Finally, run.py is the testing code which will load the python binding for object tracking.
Frankly speaking, the existences of init_wrap and process_frame_wrap come from that I don't know how to use cython to directly convert numpy array into Mat or Rect instances of OpenCv.
I will grateful if someone can show me how to do that : )