I am busy trying to build a visualisation of the by the cascalde classifier framework trained models. Up to now the FileStorage, FileNode and FileNodeIterator have helped me in grabbing every single value from the model except from the weight of each feature rectangle.
The form of a feature:
<_>
<rects>
<_>23 10 1 3 -1.</_>
<_>23 11 1 1 3.</_>
</rects>
<tilted>0</tilted></_>
<_>
Grabbing the information of each subrectangle can be done by the current code. The it_features is the iterator that runs over all features, the code handling that does just as expected and is therefore neglected.
...
vector< vector<int> > current_feature_dimensions;
FileNode rectangles = (*it_features)["rects"];
FileNodeIterator it_rect = rectangles.begin(), it_rect_end = rectangles.end();
vector<int> values;
for(int idr = 0; it_rect != it_rect_end; it_rect++, idr++ ){
(*it_rect) >> values;
current_feature_dimensions.push_back(values);
}
....
If I then visualise (only partially showed because the output is too large) this all using a loop over the vector values than the last value of the score is always 0.
for(int i = 0; i < feature_data.size(); i ++){
cerr << "Feature " << i << " - #rectangles = " << feature_data[i].size() << ": ";
for (int j = 0; j < feature_data[i].size(); j++){
vector<int> rect = feature_data[i][j];
cerr << "[" << rect[0] << " " << rect[1] << " " << rect[2] << " " << rect[3] << " " << rect[5] << "] ";
}
cerr << endl;
}
Feature 1059 - #rectangles = 2: [2 4 15 13 0] [7 4 5 13 0]
Feature 1060 - #rectangles = 2: [16 2 3 10 0] [17 2 1 10 0]
Feature 1061 - #rectangles = 2: [6 10 2 1 0] [7 10 1 1 0]
Feature 1062 - #rectangles = 2: [1 1 18 16 0] [10 1 9 16 0]
Feature 1063 - #rectangles = 2: [14 4 3 15 0] [15 4 1 15 0]
Feature 1064 - #rectangles = 2: [19 13 1 2 0] [19 14 1 1 0]
Feature 1065 - #rectangles = 2: [2 6 5 8 0] [2 10 5 4 0]
I tried the following things: changing the vector type to double and float with both reading and grabbing them from the XML file but also when writing them out with the cerr function. However this did not work. The float approach also generates a zero for the last value. The double case generates the must unrealistic values ever.
Feature 1061 - #rectangles = 2: [6 10 2 1 2.42092e-322] [7 10 1 1 2.42092e-322]
Feature 1062 - #rectangles = 2: [1 1 18 16 2.42092e-322] [10 1 9 16 2.42092e-322]
Feature 1063 - #rectangles = 2: [14 4 3 15 2.42092e-322] [15 4 1 15 2.42092e-322]
Feature 1064 - #rectangles = 2: [19 13 1 2 2.42092e-322] [19 14 1 1 2.42092e-322]
Feature 1065 - #rectangles = 2: [2 6 5 8 2.42092e-322] [2 10 5 4 2.42092e-322]
Any insights would be welcome! My debugging has lead me to believe that it is due to the fact that the FileNode interface cannot handle different types in between a single tag ... and thus this cannot be solved. On the other hand, the code base suggests that the weights is stored as a float value but in fact there is not a single float value available, they are all -1 1 -2 2 -3 3 ... so why the hell store them as floats?