Added DBScan clustering to ColorPaletteSampler

components/ColorAnalyzer/src/ColorPaletteSampler/ColorPaletteSampler.DBScan.cs

@@ -0,0 +1,156 @@
+// Licensed to the .NET Foundation under one or more agreements.
+// The .NET Foundation licenses this file to you under the MIT license.
+// See the LICENSE file in the project root for more information.
+
+using System.Numerics;
+
+namespace CommunityToolkit.WinUI.Helpers;
+
+public partial class ColorPaletteSampler
+{
+    private ref struct DBScan
+    {
+        private const int Unclassified = -1;
+
+        public static Vector3[] Cluster(Span<Vector3> points, float epsilon, int minPoints, ref float[] weights)
+        {
+            var centroids = new List<Vector3>();
+            var newWeights = new List<float>();
+
+            // Create context
+            var context = new DBScan(points, weights, epsilon, minPoints);
+
+            // Attempt to create a cluster around each point,
+            // skipping that point if already classified
+            for (int i = 0; i < points.Length; i++)
+            {
+                // Already classified, skip
+                if (context.PointClusterIds[i] is not Unclassified)
+                    continue;
+
+                // Attempt to create cluster
+                if(context.CreateCluster(i, out var centroid, out var weight))
+                {
+                    centroids.Add(centroid);
+                    newWeights.Add(weight);
+                }
+            }
+
+            weights = newWeights.ToArray();
+            return centroids.ToArray();
+        }
+
+        private bool CreateCluster(int originIndex, out Vector3 centroid, out float weight)
+        {
+            weight = 0;
+            centroid = Vector3.Zero;
+            var seeds = GetSeeds(originIndex, out bool isCore);
+
+            // Not enough seeds to be a core point.
+            // Cannot create a cluster around it
+            if (!isCore)
+            {
+                return false;
+            }
+
+            ExpandCluster(seeds, out centroid, out weight);
+            ClusterId++;
+
+            return true;
+        }
+
+        private void ExpandCluster(Queue<int> seeds,  out Vector3 centroid, out float weight)
+        {
+            weight = 0;
+            centroid = Vector3.Zero;
+            while(seeds.Count > 0)
+            {
+                var seedIndex = seeds.Dequeue();
+
+                // Skip duplicate seed entries
+                if (PointClusterIds[seedIndex] is not Unclassified)
+                    continue;
+
+                // Assign this seed's id to the cluster
+                PointClusterIds[seedIndex] = ClusterId;
+                var w = Weights[seedIndex];
+                centroid += Points[seedIndex] * w;
+                weight += w;
+
+                // Check if this seed is a core point
+                var grandSeeds = GetSeeds(seedIndex, out var seedIsCore);
+                if (!seedIsCore)
+                    continue;
+
+                // This seed is a core point. Enqueue all its seeds
+                foreach(var grandSeedIndex in grandSeeds)
+                    if (PointClusterIds[grandSeedIndex] is Unclassified)
+                        seeds.Enqueue(grandSeedIndex);
+            }
+
+            centroid /= weight;
+        }
+
+        private Queue<int> GetSeeds(int originIndex, out bool isCore)
+        {
+            var origin = Points[originIndex];
+
+            var seeds = new Queue<int>();
+            for (int i = 0; i < Points.Length; i++)
+            {
+                if (Vector3.DistanceSquared(origin, Points[i]) <= Epsilon2)
+                    seeds.Enqueue(i);
+            }
+
+            // Count includes self, so compare without checking equals
+            isCore = seeds.Count > MinPoints;
+            return seeds;
+        }
+
+        private DBScan(Span<Vector3> points, Span<float> weights, double epsilon, int minPoints)
+        {
+            Points = points;
+            Weights = weights;
+            Epsilon2 = epsilon * epsilon;
+            MinPoints = minPoints;
+
+            ClusterId = 0;
+            PointClusterIds = new int[points.Length];
+            for(int i  = 0; i < points.Length; i++)
+                PointClusterIds[i] = Unclassified;
+        }
+
+        /// <summary>
+        /// Gets the points being clustered.
+        /// </summary>
+        public Span<Vector3> Points { get; }
+
+        /// <summary>
+        /// Gets the weights of the points.
+        /// </summary>
+        public Span<float> Weights { get; }
+
+        /// <summary>
+        /// Gets or sets the id of the currently evaluating cluster.
+        /// </summary>
+        public int ClusterId { get; set; }
+
+        /// <summary>
+        /// Gets an array containing the id of the cluster each point belongs to.
+        /// </summary>
+        public int[] PointClusterIds { get; }
+
+        /// <summary>
+        /// Gets epsilon squared. Where epslion is the max distance to consider two points connected.
+        /// </summary>
+        /// <remarks>
+        /// This is cached as epsilon squared to skip a sqrt operation when comparing distances to epsilon.
+        /// </remarks>
+        public double Epsilon2 { get; }
+
+        /// <summary>
+        /// Gets the miniumum number of points required to make a core point.
+        /// </summary>
+        public int MinPoints { get; }
+    }
+}

components/ColorAnalyzer/src/ColorPaletteSampler/ColorPaletteSampler.Clustering.cs → components/ColorAnalyzer/src/ColorPaletteSampler/ColorPaletteSampler.KMeans.cs

@@ -69,7 +69,7 @@ private static void Split(int k, int[] clusterIds)
     /// <summary>
     /// Calculates the centroid of each cluster, and prunes empty clusters.
     /// </summary>
-    private static void CalculateCentroidsAndPrune(ref Span<Vector3> centroids, ref int[] counts, Span<Vector3> points, int[] clusterIds)
+    internal static void CalculateCentroidsAndPrune(ref Span<Vector3> centroids, ref int[] counts, Span<Vector3> points, int[] clusterIds)
     {
         // Clear centroids and counts before recalculation
         for (int i = 0; i < centroids.Length; i++)

components/ColorAnalyzer/src/ColorPaletteSampler/ColorPaletteSampler.cs

@@ -52,6 +52,7 @@ public async Task UpdatePaletteAsync()
 
         const int sampleCount = 4096;
         const int k = 8;
+        const float mergeDistance = 0.12f;
 
         // Retreive pixel samples from source
         var samples = await SampleSourcePixelColorsAsync(sampleCount);
@@ -62,8 +63,11 @@ public async Task UpdatePaletteAsync()
 
         // Cluster samples in RGB floating-point color space
         // With Euclidean Squared distance function, then construct palette data.
-        var clusters = KMeansCluster(samples, k, out var sizes);
-        var colorData = clusters.Select((vectorColor, i) => new PaletteColor(vectorColor.ToColor(), (float)sizes[i] / samples.Length));
+        // Merge KMeans results that are too similar, using DBScan
+        var kClusters = KMeansCluster(samples, k, out var counts);
+        var weights = counts.Select(x => (float)x / samples.Length).ToArray();
+        var dbCluster = DBScan.Cluster(kClusters, mergeDistance, 0, ref weights);
+        var colorData = dbCluster.Select((vectorColor, i) => new PaletteColor(vectorColor.ToColor(), weights[i]));
 
         // Update palettes on the UI thread
         foreach (var palette in PaletteSelectors)