Use uniform sampling as initial centers for k-means

The Wiener variance output has been sorted prior to the clustering,
which allows to directly use the uniform sampling as the initial
center points. It avoids empty cluster situations when the samples
are heavily distributed at two far ends and leave the middle empty.

Change-Id: I159fbfa6bbb4aafd19411fd005666d144cca30fc

diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 933cd3c..d6adce2 100644 (file)
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -5788,13 +5788,11 @@ int vp9_get_group_idx(double value, double *boundary_ls, int k) {
 
 void vp9_kmeans(double *ctr_ls, double *boundary_ls, int *count_ls, int k,
                 KMEANS_DATA *arr, int size) {
-  double min, max;
-  double step;
   int i, j;
   int itr;
   int group_idx;
-  double sum;
-  int count;
+  double sum[MAX_KMEANS_GROUPS];
+  int count[MAX_KMEANS_GROUPS];
 
   vpx_clear_system_state();
 
@@ -5802,38 +5800,37 @@ void vp9_kmeans(double *ctr_ls, double *boundary_ls, int *count_ls, int k,
 
   qsort(arr, size, sizeof(*arr), compare_kmeans_data);
 
-  min = arr[0].value;
-  max = arr[size - 1].value;
-
   // initialize the center points
-  step = (max - min) * 1. / k;
   for (j = 0; j < k; ++j) {
-    ctr_ls[j] = min + j * step + step / 2;
+    ctr_ls[j] = arr[(size * j) / k].value;
   }
 
   for (itr = 0; itr < 10; ++itr) {
     compute_boundary_ls(ctr_ls, k, boundary_ls);
-    group_idx = 0;
-    count = 0;
-    sum = 0;
+    for (i = 0; i < MAX_KMEANS_GROUPS; ++i) {
+      sum[i] = 0;
+      count[i] = 0;
+    }
+
     for (i = 0; i < size; ++i) {
+      // place samples into clusters
+      group_idx = 0;
       while (arr[i].value >= boundary_ls[group_idx]) {
         ++group_idx;
         if (group_idx == k - 1) {
           break;
         }
       }
+      sum[group_idx] += arr[i].value;
+      ++count[group_idx];
+    }
 
-      sum += arr[i].value;
-      ++count;
+    for (group_idx = 0; group_idx < k; ++group_idx) {
+      if (count[group_idx] > 0)
+        ctr_ls[group_idx] = sum[group_idx] / count[group_idx];
 
-      if (i + 1 == size || arr[i + 1].value >= boundary_ls[group_idx]) {
-        if (count > 0) {
-          ctr_ls[group_idx] = sum / count;
-        }
-        count = 0;
-        sum = 0;
-      }
+      sum[group_idx] = 0;
+      count[group_idx] = 0;
     }
   }