Merge pull request flann-lib#152 from pemmanuelviel/master

Better results with centers far from each other when selecting them in the chooseCentersKMeanspp stage

Author: mariusmuja

src/cpp/flann/algorithms/center_chooser.h

@@ -13,6 +13,63 @@
 namespace flann
 {
 
+template <typename Distance, typename ElementType>
+struct squareDistance
+{
+    typedef typename Distance::ResultType ResultType;
+    ResultType operator()( ResultType dist ) { return dist*dist; }
+};
+
+
+template <typename ElementType>
+struct squareDistance<L2_Simple<ElementType>, ElementType>
+{
+    typedef typename L2_Simple<ElementType>::ResultType ResultType;
+    ResultType operator()( ResultType dist ) { return dist; }
+};
+
+template <typename ElementType>
+struct squareDistance<L2_3D<ElementType>, ElementType>
+{
+    typedef typename L2_3D<ElementType>::ResultType ResultType;
+    ResultType operator()( ResultType dist ) { return dist; }
+};
+
+template <typename ElementType>
+struct squareDistance<L2<ElementType>, ElementType>
+{
+    typedef typename L2<ElementType>::ResultType ResultType;
+    ResultType operator()( ResultType dist ) { return dist; }
+};
+
+
+template <typename ElementType>
+struct squareDistance<HellingerDistance<ElementType>, ElementType>
+{
+    typedef typename HellingerDistance<ElementType>::ResultType ResultType;
+    ResultType operator()( ResultType dist ) { return dist; }
+};
+
+
+template <typename ElementType>
+struct squareDistance<ChiSquareDistance<ElementType>, ElementType>
+{
+    typedef typename ChiSquareDistance<ElementType>::ResultType ResultType;
+    ResultType operator()( ResultType dist ) { return dist; }
+};
+
+
+template <typename Distance>
+typename Distance::ResultType ensureSquareDistance( typename Distance::ResultType dist )
+{
+    typedef typename Distance::ElementType ElementType;
+
+    squareDistance<Distance, ElementType> dummy;
+    return dummy( dist );
+}
+
+
+
 template <typename Distance>
 class CenterChooser
 {
@@ -176,8 +233,11 @@ class KMeansppCenterChooser : public CenterChooser<Distance>
         assert(index >=0 && index < n);
         centers[0] = indices[index];
 
+        // Computing distance^2 will have the advantage of even higher probability further to pick new centers
+        // far from previous centers (and this complies to "k-means++: the advantages of careful seeding" article)
         for (int i = 0; i < n; i++) {
             closestDistSq[i] = distance_(points_[indices[i]], points_[indices[index]], cols_);
+            closestDistSq[i] = ensureSquareDistance<Distance>( closestDistSq[i] );
             currentPot += closestDistSq[i];
         }
 
@@ -203,7 +263,10 @@ class KMeansppCenterChooser : public CenterChooser<Distance>
 
                 // Compute the new potential
                 double newPot = 0;
-                for (int i = 0; i < n; i++) newPot += std::min( distance_(points_[indices[i]], points_[indices[index]], cols_), closestDistSq[i] );
+                for (int i = 0; i < n; i++) {
+                    DistanceType dist = distance_(points_[indices[i]], points_[indices[index]], cols_);
+                    newPot += std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
+                }
 
                 // Store the best result
                 if ((bestNewPot < 0)||(newPot < bestNewPot)) {
@@ -215,7 +278,10 @@ class KMeansppCenterChooser : public CenterChooser<Distance>
             // Add the appropriate center
             centers[centerCount] = indices[bestNewIndex];
             currentPot = bestNewPot;
-            for (int i = 0; i < n; i++) closestDistSq[i] = std::min( distance_(points_[indices[i]], points_[indices[bestNewIndex]], cols_), closestDistSq[i] );
+            for (int i = 0; i < n; i++) {
+                DistanceType dist = distance_(points_[indices[i]], points_[indices[bestNewIndex]], cols_);
+                closestDistSq[i] = std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
+            }
         }
 
         centers_length = centerCount;