logistic_regression.m

function logistic_regression(is_train, f_p, f_n)
% train the pairwise model
% Input1: positive features: f_p
% Input2: negative features: f_n
% To Jie: you need to download liblinear to learn the logistic_regression
addpath(genpath('/BS/pedestrian-detection-tracking/work/project/3rd/liblinear-1.94/'));

if is_train == 1 % train 
    num_positive = size(f_p,1);
    num_negative = size(f_n,1);
    
    % normalize the features;
    % To jie: f_p is the feature list for the positive examples (e.g. part field evalue of pair of detections from the same person), f_n is the feature list from the negative examples
    % To jie:  the features have to be normalized to 0~1
    [f_normalize, normalize_ratio]= normalize_feature([f_p; f_n]);
    f_p = f_normalize(1:num_positive,:);
    f_n = f_normalize(1+num_positive:num_positive+num_negative,:);
    
    
    label = [ones(num_positive,1); zeros(num_negative,1)];
    example = sparse(double([f_p;f_n]));
    
    % L2 regularizer and C = 1; one can do cross validation for c
    % parameters
    regularizer_para = 0;
    c_para = 1;
    
    % To jie: call the train function in liblinear to obtain the model
    model = train(label, example, ['-s ' num2str(regularizer_para) ' -B ' num2str(1) ' -c ' num2str(c_para) ]);
    
    model_name = 'first_model.mat';
    save(model_name , 'model', 'normalize_ratio');
    
else %test
    load('first_model.mat');
    
    f = f_p;
    f =  normalize_feature(f, normalize_ratio);
    
    f = sparse(double(f));
    label = ones(size(f,1),1);
    [~, ~, pro_estimate] = predict(label, f, model,  '-b 1 -q');
    
    probablity = pro_estimate(:,1);
end
end