Oswaldo Ludwig

Mahesh,

Regarding the application of my codes in multi-class problems, the feature selector based on genetic algorithms and information theory can deal with multi-class problem by adopting the target data, y, as a line vector in which each class is represented by a natural number, such as 1 for the first class, 2 for the second... The SVNN can be upgraded to solve multi-class problem by adapting the framework described in "K. Crammer and Y. Singer. On the Algorithmic Implementation of Multi-class SVMs, JMLR, 2001". However, since I finished my PhD, I have no time to work on this research line.

An application for human detection (Windows only) can be found in http://webmail.isr.uc.pt/~oludwig/
The program is based on HOG descriptors for Matlab arranged in SRM-cascade, according to the methodology described in: O. Ludwig; “ Study on Non-parametric Methods for Fast Pattern Recognition with Emphasis on Neural Networks and Cascade Classifiers;” PhD Thesis, University of Coimbra, Coimbra, 2012. New images can be included in the folder "images" with the name "ped_XX". The algorithm was trained by using 18x36 gray-scale images from the Daimler Pedestrian Classification dataset. Double-click in detector_pkg.exe to install the program, then it is possible to run detector.exe.

Dear Mr Ludwig,

have to thanks you for such Great work, I did not understand it 100% so just could not use it,
I have a multi-class problem, I do solve it with "classic" MLP, but i dont know how to use you functions,
so please, if u can send an exemple, you or anyone else reading this, it would realy Help a lot.

I noticed a bug in the histogram computation. Gradient vectors whose angle is equal to pi are being omitted from the histogram because you check for angle < angle_lim instead of angle <= angle_lim.

I also re-wrote the histogram computation to optimize it in a way similar to what was suggested in an earlier comment.

% Compute the bin index for every angle.
binIndeces = ceil((v_angles + pi) / binSize);

% For each bin
for (bin = 1:B)
% Get the magnitudes for all the pixels that belong in 'bin'.
magnitudes = v_magnit(binIndeces == bin);

% Add all of the magnitudes to 'bin'.
H2(bin) = H2(bin) + sum(magnitudes);
end

Could you explain why it was adopted blockwise normalization? From the best of my knowledge, there is no mathematical foundation behind HOG descriptors, only good sense. This is my version, which was made available to be evaluated by the community. Go ahead, you should make the comparison with other algorithms in your case study (Science is not religion, feel free to doubt and inovate).