Added an additional threshold for each detection template. This is useful

since some detection templates may be larger than others and therefore
naturally accumulate larger scores, therefore necessitating larger threshold
values.

dlib/image_processing/scan_image_pyramid.h

@@ -176,7 +176,8 @@ namespace dlib
             rectangle& mapped_rect,
             detection_template& best_template,
             rectangle& object_box,
-            unsigned long& best_level
+            unsigned long& best_level,
+            unsigned long& detection_template_idx
         ) const;
 
         double get_match_score (
@@ -580,7 +581,7 @@ namespace dlib
             << "\n\t this: " << this
             );
 
-        return feats_config.get_num_dimensions()*get_num_components_per_detection_template();
+        return feats_config.get_num_dimensions()*get_num_components_per_detection_template() + get_num_detection_templates();
     }
 
 // ----------------------------------------------------------------------------------------
@@ -661,7 +662,7 @@ namespace dlib
             for (unsigned long i = 0; i < saliency_images.size(); ++i)
             {
                 saliency_images[i].set_size(feats[l].nr(), feats[l].nc());
-                const unsigned long offset = feats_config.get_num_dimensions()*i;
+                const unsigned long offset = get_num_detection_templates() + feats_config.get_num_dimensions()*i;
 
                 // build saliency images for pyramid level l 
                 for (long r = 0; r < feats[l].nr(); ++r)
@@ -702,14 +703,21 @@ namespace dlib
                 rectangle scaled_object_box = feats[l].image_to_feat_space(det_templates[i].object_box);
                 scaled_object_box = centered_rect(point(0,0),scaled_object_box.width(), scaled_object_box.height());
 
+                // Each detection template gets its own special threshold in addition to
+                // the global detection threshold.  This allows us to model the fact that
+                // some detection templates might be more prone to false alarming or since
+                // their size is different naturally require a larger or smaller threshold
+                // (since they integrate over a larger or smaller region of the image).
+                const double template_specific_thresh = w(i);
+
                 scan_image_movable_parts(point_dets, saliency_images, scaled_object_box,
                                          stationary_region_rects, movable_region_rects,
-                                         thresh, max_dets_per_template); 
+                                         thresh+template_specific_thresh, max_dets_per_template); 
 
                 // convert all the point detections into rectangles at the original image scale and coordinate system
                 for (unsigned long j = 0; j < point_dets.size(); ++j)
                 {
-                    const double score = point_dets[j].first;
+                    const double score = point_dets[j].first+template_specific_thresh;
                     point p = point_dets[j].second;
                     p = feats[l].feat_to_image_space(p);
                     rectangle rect = translate_rect(det_templates[i].object_box, p);
@@ -744,8 +752,8 @@ namespace dlib
 
         rectangle mapped_rect, object_box;
         detection_template best_template;
-        unsigned long best_level;
-        get_mapped_rect_and_metadata(max_pyramid_levels, rect, mapped_rect, best_template, object_box, best_level);
+        unsigned long best_level, junk;
+        get_mapped_rect_and_metadata(max_pyramid_levels, rect, mapped_rect, best_template, object_box, best_level, junk);
         return mapped_rect;
     }
 
@@ -762,7 +770,8 @@ namespace dlib
         rectangle& mapped_rect,
         detection_template& best_template,
         rectangle& object_box,
-        unsigned long& best_level
+        unsigned long& best_level,
+        unsigned long& detection_template_idx
     ) const
     {
         pyramid_type pyr;
@@ -799,6 +808,7 @@ namespace dlib
                     best_match_score = match_score;
                     best_level = l;
                     best_template = det_templates[t];
+                    detection_template_idx = t;
                 }
             }
         }
@@ -849,9 +859,9 @@ namespace dlib
 
         rectangle mapped_rect;
         detection_template best_template;
-        unsigned long best_level;
+        unsigned long best_level, junk;
         rectangle object_box;
-        get_mapped_rect_and_metadata(feats.size(), rect, mapped_rect, best_template, object_box, best_level);
+        get_mapped_rect_and_metadata(feats.size(), rect, mapped_rect, best_template, object_box, best_level, junk);
 
         Pyramid_type pyr;
 
@@ -978,9 +988,11 @@ namespace dlib
 
         rectangle mapped_rect;
         detection_template best_template;
-        unsigned long best_level;
+        unsigned long best_level, detection_template_idx;
         rectangle object_box;
-        get_mapped_rect_and_metadata (feats.size(), obj.get_rect(), mapped_rect, best_template, object_box, best_level);
+        get_mapped_rect_and_metadata(feats.size(), obj.get_rect(), mapped_rect, best_template, object_box, best_level, detection_template_idx);
+
+        psi(detection_template_idx) += 1;
 
         Pyramid_type pyr;
 
@@ -1010,7 +1022,7 @@ namespace dlib
         {
             const rectangle rect = rects[j].intersect(get_rect(feats[best_level]));
             const unsigned long template_region_id = j;
-            const unsigned long offset = feats_config.get_num_dimensions()*template_region_id;
+            const unsigned long offset = get_num_detection_templates() + feats_config.get_num_dimensions()*template_region_id;
             for (long r = rect.top(); r <= rect.bottom(); ++r)
             {
                 for (long c = rect.left(); c <= rect.right(); ++c)