Detection from matlab and c++ different

[adtadhikari]

I tried running the detection in c++ and matlab but it is not the same. The matlab detection is better than the c++ one. What may be the reason? What are the difference between the matlab and the c++ code? I am using the Person_26parts model.

[adtadhikari]

@hbristow do you have any idea regarding this? What is the difference in the human pose detection for 26 parts between matlab and the c++ version?

[AASanchezA]

I found a similar issue, I tried the original code from Ramanan (http://www.ics.uci.edu/~dramanan/software/pose/pose-release1.2-full.tgz) and then tried your code, and the result are similar, but there is displacement between the result, I checked the visualization part and is not there the problem, my idea is that it could be where you generate the boundingbox, because the upper left corner of the boundingbox coincides with almost the center bounding box in the matlab implementation, this pics show the problem,

C++

matlab Ramanan

skeleton best

[adtadhikari]

Yeah my problem was similar but I was testing it with people. Please let me know if you are able to solve this problem.

[AASanchezA]

HI, I didn't propperly solve the problem, but I found where is it, is in the DynamicProgram.cpp, where you calculate the bounding rectangle and add it to the Candidate, thre is a rounding problem between the scale and the size of the image, my solution in not working properly, but for my it is enough!! you can check my fork

[legolas123]

I am also seeing a huge difference in the outputs between matlab and c++ code. Did anyone find difference between these sets of codes?

[kvikrama]

In Matlab code pyramid generation also creates a border. This part of code is commented in C code.
Uncommenting also didnt improve the detection.
I feel that scaling (resizing) is different in C++ and matlab. Has anyone made progress in this area ?
I am also struck here for sometime.

[simonamaggio]

Hi all,

I also found out that the proposed xml model for people (Person_26parts.xml) is slightly different from the matlab model PARSE_model.mat released in http://www.ics.uci.edu/~dramanan/software/pose/pose-release1.2-full.tgz. Generating a new .xml file starting from the .mat models resulted in a more precise person detector.