Merge pull request #2016 from pemmanuelviel:kmeansppSquareDist

modules/flann/include/opencv2/flann/dist.h

@@ -812,6 +812,66 @@ struct ZeroIterator
 
 };
 
+
+/*
+ * Depending on processed distances, some of them are already squared (e.g. L2)
+ * and some are not (e.g.Hamming). In KMeans++ for instance we want to be sure
+ * we are working on ^2 distances, thus following templates to ensure that.
+ */
+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<ElementType>, ElementType>
+{
+    typedef typename L2<ElementType>::ResultType ResultType;
+    ResultType operator()( ResultType dist ) { return dist; }
+};
+
+
+template <typename ElementType>
+struct squareDistance<MinkowskiDistance<ElementType>, ElementType>
+{
+    typedef typename MinkowskiDistance<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 );
+}
+
 }
 
 #endif //OPENCV_FLANN_DIST_H_

modules/flann/include/opencv2/flann/hierarchical_clustering_index.h

@@ -210,8 +210,11 @@ private:
         assert(index >=0 && index < n);
         centers[0] = dsindices[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(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
+            closestDistSq[i] = ensureSquareDistance<Distance>( closestDistSq[i] );
             currentPot += closestDistSq[i];
         }
 
@@ -237,7 +240,10 @@ private:
 
                 // Compute the new potential
                 double newPot = 0;
-                for (int i = 0; i < n; i++) newPot += std::min( distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols), closestDistSq[i] );
+                for (int i = 0; i < n; i++) {
+                    DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[index]], dataset.cols);
+                    newPot += std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
+                }
 
                 // Store the best result
                 if ((bestNewPot < 0)||(newPot < bestNewPot)) {
@@ -249,7 +255,10 @@ private:
             // Add the appropriate center
             centers[centerCount] = dsindices[bestNewIndex];
             currentPot = bestNewPot;
-            for (int i = 0; i < n; i++) closestDistSq[i] = std::min( distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols), closestDistSq[i] );
+            for (int i = 0; i < n; i++) {
+                DistanceType dist = distance(dataset[dsindices[i]], dataset[dsindices[bestNewIndex]], dataset.cols);
+                closestDistSq[i] = std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
+            }
         }
 
         centers_length = centerCount;

modules/flann/include/opencv2/flann/kmeans_index.h

@@ -211,6 +211,7 @@ public:
 
         for (int i = 0; i < n; i++) {
             closestDistSq[i] = distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols);
+            closestDistSq[i] = ensureSquareDistance<Distance>( closestDistSq[i] );
             currentPot += closestDistSq[i];
         }
 
@@ -236,7 +237,10 @@ public:
 
                 // Compute the new potential
                 double newPot = 0;
-                for (int i = 0; i < n; i++) newPot += std::min( distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols), closestDistSq[i] );
+                for (int i = 0; i < n; i++) {
+                    DistanceType dist = distance_(dataset_[indices[i]], dataset_[indices[index]], dataset_.cols);
+                    newPot += std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
+                }
 
                 // Store the best result
                 if ((bestNewPot < 0)||(newPot < bestNewPot)) {
@@ -248,7 +252,10 @@ public:
             // Add the appropriate center
             centers[centerCount] = indices[bestNewIndex];
             currentPot = bestNewPot;
-            for (int i = 0; i < n; i++) closestDistSq[i] = std::min( distance_(dataset_[indices[i]], dataset_[indices[bestNewIndex]], dataset_.cols), closestDistSq[i] );
+            for (int i = 0; i < n; i++) {
+                DistanceType dist = distance_(dataset_[indices[i]], dataset_[indices[bestNewIndex]], dataset_.cols);
+                closestDistSq[i] = std::min( ensureSquareDistance<Distance>(dist), closestDistSq[i] );
+            }
         }
 
         centers_length = centerCount;