CODEWITHSUNDEEP
c++image-processingmachine-learningface-detectiondlib
Icarium
Icarium

Reputation: 73

train dlib shape predictor with 194 points helen dataset

I have an issue with preparing shape training with dlib library (train_shape_predictor_ex.cpp) To the case:

I properly installed Visual studio and configured dlib properly Also I have xml file with 194 points landmarks from helen database (about 2300 images mapped in total).

By default the dlib solution is based on face data set that are related to 68 face landmarks and I simply don't know how to expand the limitation from 68 to 194.

When I tried the best what I got that far are 68 points mapped on 194 points data set per image...

By reverse engineering I have found that in render_face_detections.h file there are limitations but even when I changed them from 68 to 194 I have still same output which displays only 68 landmark point on the analyzed image.

I would be grateful for any hints how to solve this problem - I know it's a simple adjustment but I couldn't figure it out..

Thanks in advance, Bartek

Upvotes: 1

Views: 999

Answers (1)

ManuelTS
ManuelTS

Reputation: 356

Maybe because you don't know the structure of the 194 points and how they are alinged. I would suggest using only circles as I do:

inline std::vector<image_window::overlay_circle> render_helen_face_detections (
        const std::vector<full_object_detection>& dets,
        const rgb_pixel color = rgb_pixel(0,255,0)
    )
    {
        std::vector<image_window::overlay_circle> circles;

        for (unsigned long i = 0; i < dets.size(); i++){
          const full_object_detection& d = dets[i];

          for (unsigned long p = 0; p < d.num_parts(); p++)
              circles.push_back(image_window::overlay_circle(d.part(p), 1, color));
        }

        return circles;
    }

Upvotes: 1

Related Questions