OpenCv 3 Beta -ICF classifier performance is very bad.

edit

Hi,

I am working with OpenCV_Contrib module and using xobjdetect module for pedestrian detection. I have trained ICF model using following params:

icfParams.alpha= 0.02;
icfParams.bg_per_image= 5;
icfParams.feature_count= 30000;
icfParams.features_type= "icf";
icfParams.is_grayscale= false ;
icfParams.model_n_cols= 64;  // W-->Training Model Size ..Size of +Ve and -Ve image
icfParams.model_n_rows= 128; // H-->Training Model Size ..Size of +Ve and -Ve image
icfParams.use_fast_log= false;
icfParams.weak_count= 2000;

with Inrai data set

#positive image =2416, # no of neg=1218 ( total icfParams.bg_per_image *1218);

For this I have been using the following software

vector<cv::Rect> objects;
vector<float> values;
float scaleFactor=1.09; // 2^1/8;
float threshold=0;
cv::Size minSize= cv::Size(24,48); //(width,height)
cv::Size maxSize= cv::Size(64,128);
cout <<minSize.height << minSize.width <<endl;
int slidingStep=8;
icfDetect.detect(img,objects,scaleFactor,minSize, maxSize, threshold, slidingStep,values);

I am trying to detect Pedestrian in given Image.

INPUT IMAGE

OUTPUT IMAGE

Plz someone suggest what is the issue? Also, why this using waldboost+ stump instead of 3 level decision tree as suggested by author Dollar in his paper at http://vision.ucsd.edu/sites/default/...

Few doubts-

• How to train model in better ways.
• How to get better detection rate/window
• How to set these params.