Revamp face detection to reduce false positives.

Plugins/Faces/FaceDetection.cs

@@ -62,14 +62,20 @@ public class FaceDetection:FeatureDetectionBase<Face>{
         public FaceDetection():base(){
             MinFaces = 1;
             MaxFaces = 10;
-            MinSizePercent = 3;
-            ConfidenceLevelThreshold = 2;
-            MinConfidenceLevel = 1;
+            MinSizePercent = 4;
+            ConfidenceLevelThreshold = 5;
+            MinConfidenceLevel = 3;
 
             ExpandX = 0;
             ExpandY = 0;
             fileNames = new Dictionary<string, string>(){ 
-            {"FaceCascade",@"haarcascade_frontalface_default.xml"}, {"FaceCascadeAlt",@"haarcascade_frontalface_alt.xml"} };
+            {"FaceCascade",@"haarcascade_frontalface_default.xml"},
+                { "FaceCascadeAlt",@"haarcascade_frontalface_alt.xml"},
+                { "FaceCascadeAlt2",@"haarcascade_frontalface_alt2.xml"},
+                {"FaceCascadeAltTree",@"haarcascade_frontalface_alt_tree.xml"},
+                {"FaceProfile",@"haarcascade_profileface.xml"},
+                {"Eye",@"haarcascade_eye.xml"},
+            };
         }
         /// <summary>
         /// The minimum number of faces expected
@@ -106,8 +112,9 @@ public class FaceDetection:FeatureDetectionBase<Face>{
         /// The percentage by which to expand each face rectangle vertically after detection. To expand 20% on the top and bottom, set to 0.4
         /// </summary>
         public double ExpandY { get; set; }
-
-
+
+        static long totalTime = 0;
+        static long count = 0;
         /// <summary>
         /// Detects features on a grayscale image.
         /// </summary>
@@ -117,28 +124,59 @@ public class FaceDetection:FeatureDetectionBase<Face>{
         protected override List<Face> DetectFeatures(IplImage img, CvMemStorage storage) {
 
             //Determine minimum face size
-            var minSize = (int)Math.Round((double)MinSizePercent / 100.0 * Math.Min(img.Width, img.Height));
+            var minSize = Math.Max(12, (int)Math.Round((double)MinSizePercent / 100.0 * Math.Min(img.Width, img.Height)));
 
 
-            //Detect faces (frontal). TODO: side
+            //Detect faces (frontal). 
             Stopwatch watch = Stopwatch.StartNew();
-            // HaarDetectionType.DoCannyPruning | HaarDetectionType.ScaleImage
-
-            CvAvgComp[] faces = BorrowCascade("FaceCascade", c => Cv.HaarDetectObjects(img, c, storage, 1.0850, MinConfidenceLevel, HaarDetectionType.Zero, new CvSize(minSize, minSize), new CvSize(0,0)).ToArrayAndDispose());
-            watch.Stop();
-            Debug.WriteLine($"Face detection time: {watch.ElapsedMilliseconds}ms");
+
 
+            CvAvgComp[] faces = BorrowCascade("FaceCascadeAlt", c => Cv.HaarDetectObjects(img, c, storage, 1.0850, MinConfidenceLevel, HaarDetectionType.DoCannyPruning, new CvSize(minSize, minSize), new CvSize(0,0)).ToArrayAndDispose());
+
             //Sort by accuracy
             Array.Sort<CvAvgComp>(faces, CompareByNeighbors);
 
             //Convert into feature objects list
             List<Face> features = new List<Face>(faces.Length);
             foreach (CvAvgComp face in faces) features.Add(new Face(PolygonMath.ScaleRect(face.Rect.ToRectangleF(),ExpandX,ExpandY), face.Neighbors));
+
+            // Doesn't add much, and would have to be deduplicated.
+            //CvAvgComp[] profiles = BorrowCascade("FaceProfile", c => Cv.HaarDetectObjects(img, c, storage, 1.2, MinConfidenceLevel + 2, HaarDetectionType.FindBiggestObject | HaarDetectionType.DoRoughSearch | HaarDetectionType.DoCannyPruning, new CvSize(img.Width / 8, img.Height / 8), new CvSize(0, 0)).ToArrayAndDispose());
+            //foreach (CvAvgComp face in profiles) features.Add(new Face(PolygonMath.ScaleRect(face.Rect.ToRectangleF(), ExpandX, ExpandY), face.Neighbors));
+
+
+            // Test for eyes, if faces > 20 pixels
+            foreach (var face in features) {
+                var w = (int) (face.X2 - face.X);
+                var h = (int) ((face.Y2 - face.Y) * 0.6);
+                if (w > 20) {
+                    img.SetROI((int) face.X, (int) face.Y, w, h);
+                    storage.Clear(); 
+                    CvAvgComp[] eyes = BorrowCascade("Eye",
+                        c => Cv.HaarDetectObjects(img, c, storage, 1.0850, 4, HaarDetectionType.FindBiggestObject | HaarDetectionType.DoRoughSearch,
+                                   new CvSize(4, 4), new CvSize(img.Width / 2, img.Height / 2))
+                               .ToArrayAndDispose());
+                    if (eyes.Length == 0) {
+                        // Halve the estimated accuracy if there are no eyes detected
+                        face.Accuracy = face.Accuracy / 2;
+                        // We never want to boost accuracy, because the walls have eyes
+                    }
+                }
+            }
+
 
+
+
             //Unless we're below MinFaces, filter out the low confidence matches.
             while (features.Count > MinFaces && features[features.Count - 1].Accuracy < ConfidenceLevelThreshold) features.RemoveAt(features.Count - 1);
 
 
+            watch.Stop();
+            totalTime += watch.ElapsedMilliseconds;
+            count++;
+            Debug.WriteLine($"Face detection time: {watch.ElapsedMilliseconds}ms  (avg {totalTime / count}ms)");
+
+
             //Never return more than [MaxFaces]
             return (features.Count > MaxFaces) ? features.GetRange(0, MaxFaces) : features;
         }

Plugins/Faces/FacesPlugin.cs

@@ -140,7 +140,10 @@ public List<Face> GetFacesFromImage(object image, NameValueCollection settings)
 
                     g.DrawRectangle(Pens.Green, new Rectangle((int)x1, (int)y1, (int)(x2 - x1), (int)(y2 - y1)));
                 }
+
+                g.DrawString((newPoints.Length / 2).ToString(), new Font(FontFamily.GenericSansSerif, 24.0f, FontStyle.Bold), new SolidBrush(Color.Green),0.0f,24.0f);
             }
+
 
 
             return RequestedAction.None;
@@ -187,10 +190,7 @@ public List<Face> GetFacesFromImage(object image, NameValueCollection settings)
 
             //Parse min/max faces
             int[] count = s.GetList<int>("f.faces",null,1,2);
-            if (count == null) {
-                f.MinFaces = 1;
-                f.MaxFaces = 8;
-            }else if (count.Length > 0){
+            if (count != null && count.Length > 0){
                 f.MinFaces = f.MaxFaces = count[0];
                 if (count.Length > 1) f.MaxFaces = count[1];
             }

Plugins/Faces/FeatureDetectionBase.cs

@@ -81,7 +81,7 @@ protected TR BorrowCascade<TR>(string fileNameKey, Func<CvHaarClassifierCascade,
         /// Large images will be scaled down to less than scaledBounds X scaledBounds for feature detection.
         /// Defaults to 1000
         /// </summary>
-        protected int scaledBounds = 1000;
+        protected int scaledBounds = 800;
 
         public List<T> DetectFeatures(Bitmap b)
         {
@@ -92,16 +92,20 @@ public List<T> DetectFeatures(Bitmap b)
 
             IplImage orig = null;
             IplImage gray = null;
+            IplImage gray2 = null;
             IplImage small = null;
             try {
 
                 orig = OpenCvSharp.Extensions.BitmapConverter.ToIplImage(b);
 
 
                 gray = new IplImage(orig.Size, BitDepth.U8, 1);
+                //gray2 = new IplImage(orig.Size, BitDepth.U8, 1);
+
                 //Make grayscale version
                 Cv.CvtColor(orig, gray, ColorConversion.BgrToGray); //TODO, try a different color space
-
+                //Cv.EqualizeHist(gray, gray2);
+
                 var w = orig.Width;
                 var h = orig.Height;
                 Cv.ReleaseImage(orig);
@@ -140,6 +144,7 @@ public List<T> DetectFeatures(Bitmap b)
                 }
             } finally {
                 if (gray != null) Cv.ReleaseImage(gray);
+                if (gray2 != null) Cv.ReleaseImage(gray2);
                 if (orig != null) Cv.ReleaseImage(orig);
                 if (small != null) Cv.ReleaseImage(small);
             }

Plugins/Faces/Native.xml

@@ -1,8 +1,8 @@
 <?xml version="1.0" encoding="utf-8" ?>
 <files>
-  <file url="https://d3ndcb4i803ljg.cloudfront.net/opencv/2.4.10/cascades/haarcascade_frontalface_default.xml" name="haarcascade_frontalface_default.xml"  fileBytes="1254733" />
 
   <file url="https://d3ndcb4i803ljg.cloudfront.net/opencv/2.4.10/cascades/haarcascade_frontalface_alt.xml" name="haarcascade_frontalface_alt.xml"  fileBytes="919871" />
+  <file url="https://d3ndcb4i803ljg.cloudfront.net/opencv/2.4.10/cascades/haarcascade_eye.xml" name="haarcascade_eye.xml"  fileBytes="506314" />
 
 
   <file url="https://d3ndcb4i803ljg.cloudfront.net/opencv/2.4.10/x86/opencv_legacy2410.dll" name="opencv_legacy2410.dll" bitness="32" fileBytes="1239040" />