Added embedding extractor. DeepSeek next.

AGENTS.md

@@ -13,7 +13,7 @@ Splitter pipeline is:
   * FFMpeg used to decode video frames into OpenCVSharp.Mat
   * One of detectors used:
     - For face detection: [opencv_zoo/models/face_detection_yunet at main opencv/opencv_zoo](https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet)
-    - For body detection: [yolov8s.pt Ultralytics/YOLOv8 at main](https://huggingface.co/Ultralytics/YOLOv8/blob/main/yolov8s.pt)
+    - For body detection: [THU-MIG/yolov10: YOLOv10: Real-Time End-to-End Object Detection [NeurIPS 2024]](https://github.com/THU-MIG/yolov10/tree/main)
   * Camera control aplied (CameraControl class)
   * Final video frames are encoded back to video file using FFMpeg
 

Splitter-UI/Models/PreviewData.cs

@@ -6,13 +6,17 @@ public class PreviewData
     public IReadOnlyList<OpenCvSharp.Rect> DetectedBoxes { get; }
     public Rect? CropRect                                { get; }
     public Point2f GravitateTo                           { get; }
+    public TimeSpan Position                             { get; }
+    public int? Rotate                                   { get; }
 
-    public PreviewData(Avalonia.Media.Imaging.Bitmap? frame, IReadOnlyList<OpenCvSharp.Rect> boxes, Rect? crop, Point2f gravitateTo)
+    public PreviewData(Avalonia.Media.Imaging.Bitmap? frame, IReadOnlyList<OpenCvSharp.Rect> boxes, Rect? crop, Point2f gravitateTo, TimeSpan position, int? rotate)
     {
         Frame         = frame;
         DetectedBoxes = boxes;
         CropRect      = crop;
         GravitateTo   = gravitateTo;
+        Position      = position;
+        Rotate        = rotate;
     }
 
 }

Splitter-UI/Program.cs

@@ -36,16 +36,19 @@ internal sealed class Program
 
         // splitter services
         services.AddSingleton<UltraFaceDetector>();
-        services.AddSingleton<YoloOnnxObjectDetector>();
-        services.AddSingleton( x => new SingleThreadedDetector<UltraFaceDetector>(x.GetRequiredService<UltraFaceDetector>()) );
-        services.AddSingleton(x => new SingleThreadedDetector<YoloOnnxObjectDetector>(x.GetRequiredService<YoloOnnxObjectDetector>()));
+        services.AddSingleton<YoloV10ObjectDetector>();
+        services.AddSingleton<OSNetEmbeddingExtractor>();
+        services.AddSingleton<IObjectTracker, ObjectTracker>();
+        services.AddSingleton(x => new SingleThreadedDetector<UltraFaceDetector>(x.GetRequiredService<UltraFaceDetector>()));
+        services.AddSingleton(x => new SingleThreadedDetector<YoloV10ObjectDetector>(x.GetRequiredService<YoloV10ObjectDetector>()));
         services.AddSingleton(x => new SingleThreadedDetector<DummyDetector>(x.GetRequiredService<DummyDetector>()));
+        services.AddSingleton<IEmbeddingExtractor>(x => new SingleThreadedEmbeddingExtractor<OSNetEmbeddingExtractor>(x.GetRequiredService<OSNetEmbeddingExtractor>()));
         services.AddSingleton<Func<string, IObjectDetector>>( x => detectorName =>
         {
             return detectorName switch
             {
                 "face" => x.GetRequiredService<SingleThreadedDetector<UltraFaceDetector>>(),
-                "body" => x.GetRequiredService<SingleThreadedDetector<YoloOnnxObjectDetector>>(),
+                "body" => x.GetRequiredService<SingleThreadedDetector<YoloV10ObjectDetector>>(),
                 "none" => x.GetRequiredService<SingleThreadedDetector<DummyDetector>>(),
                 _ => new DummyDetector()
             };

Splitter-UI/Services/SingleThreadedDetector.cs

@@ -5,7 +5,7 @@ public class SingleThreadedDetector<T>(IObjectDetector _detector) : IObjectDetec
 {
     private Lock _lock = new();
 
-    public List<(OpenCvSharp.Rect box, Point2f center)> DetectAll(SingleTask job, Mat frameCont)
+    public List<DetectedPerson> DetectAll(SingleTask job, Mat frameCont)
     {
         lock (_lock)
         {
@@ -19,3 +19,24 @@ public class SingleThreadedDetector<T>(IObjectDetector _detector) : IObjectDetec
             d.Dispose();
     }
 }
+
+public class SingleThreadedEmbeddingExtractor<T>(IEmbeddingExtractor _extractor) : IEmbeddingExtractor
+    where T : IEmbeddingExtractor
+{
+    private Lock _lock = new();
+
+    public float[] Extract(Mat frame, OpenCvSharp.Rect box)
+    {
+        lock (_lock)
+        {
+            return _extractor.Extract(frame, box);
+        }
+    }
+
+    public void Dispose()
+    {
+        if (_extractor is IDisposable d)
+            d.Dispose();
+    }
+
+}

Splitter-UI/ViewModels/JobViewModel.cs

@@ -15,7 +15,7 @@ public partial class JobViewModel : ObservableObject
     public SingleJob GetJob() => Job;
 
     [ObservableProperty] private VideoInfo?    _probe;
-    [ObservableProperty] private PreviewData?  _preview = new(null, [], null, new(0.5f, 0.5f));
+    [ObservableProperty] private PreviewData?  _preview = new(null, [], null, new(0.5f, 0.5f), TimeSpan.Zero, null);
     [ObservableProperty] private Bitmap?       _thumbnail;
     [ObservableProperty] private double        _sliderLiveValue;
     [ObservableProperty] private double        _positionSeconds;
@@ -70,7 +70,7 @@ public partial class JobViewModel : ObservableObject
         {
             if (string.IsNullOrWhiteSpace(value))
             {
-                Job.GravitateTo = null;
+                Job.GravitateTo = new Point2f(0.5f, 0.5f);
             }
             else
             {
@@ -109,6 +109,19 @@ public partial class JobViewModel : ObservableObject
         }
     }
 
+    public float ScoreThreshold
+    {
+        get => Job.ScoreThreshold;
+        set
+        {
+            if (Math.Abs(Job.ScoreThreshold - value) < 0.001)
+                return;
+            Job.ScoreThreshold = value;
+            OnPropertyChanged();
+            Task.Run(CreatePreview);
+        }
+    }
+
     public string? Mask
     {
         get => Job.Mask;
@@ -182,10 +195,10 @@ public partial class JobViewModel : ObservableObject
 
     public Point2f GravitateTo
     {
-        get => Job.GravitateTo ?? new Point2f(0.5f, 0.5f);
+        get => Job.GravitateTo;
         set
         {
-            if (Job.GravitateTo != null && Math.Abs(Job.GravitateTo.Value.X - value.X) < 0.001 && Math.Abs(Job.GravitateTo.Value.Y - value.Y) < 0.001)
+            if (Math.Abs(Job.GravitateTo.X - value.X) < 0.001 && Math.Abs(Job.GravitateTo.Y - value.Y) < 0.001)
                 return;
 
             Job.GravitateTo = value;
@@ -199,6 +212,8 @@ public partial class JobViewModel : ObservableObject
         get => Job.DetectAbove;
         set
         {
+            if (Math.Abs(Job.DetectAbove - value) < 0.001 )
+                return;
             Job.DetectAbove = value;
             OnPropertyChanged();
             Task.Run(CreatePreview);
@@ -262,11 +277,17 @@ public partial class JobViewModel : ObservableObject
             return;
         try
         {
-            var frame    = await _thumbnails.CreateThumbnailAsync(Job.InputFile, Probe, TimeSpan.FromSeconds(PositionSeconds), Probe.Width, Probe.Height, Job.Rotate);
+            var pos = TimeSpan.FromSeconds(PositionSeconds);
+
+            Bitmap? frame;
+            if (Preview?.Frame == null || Preview.Position != pos) 
+                frame = await _thumbnails.CreateThumbnailAsync(Job.InputFile, Probe, pos, Probe.Width, Probe.Height, Job.Rotate);
+            else
+                frame = Preview.Frame;
             if (  frame == null )
                 return;
 
-            Preview      = new PreviewData(frame, [], null, Job.GravitateTo ?? new (0.5f, 0.5f));
+            Preview      = new PreviewData(frame, [], null, Job.GravitateTo, pos, Job.Rotate);
 
             var detector = _detectorFactory(Job.Detect ?? "");
             var j = new SingleTask
@@ -286,7 +307,7 @@ public partial class JobViewModel : ObservableObject
             if (detections.Count > 0)
             {
                 var primaryDetection = detections
-                    .OrderByDescending(d => d.box.Height * d.box.Width)
+                    .OrderByDescending(d => d.Box.Height * d.Box.Width)
                     .FirstOrDefault();
 
                 var w = Probe.Width;
@@ -295,16 +316,16 @@ public partial class JobViewModel : ObservableObject
                 var cropWidth  = Job.Crop?.width  ?? CommandLine.DefaultW;
                 var cropHeight = Job.Crop?.height ?? CommandLine.DefaultH;
 
-                var cx = primaryDetection.center.X - cropWidth  / 2f;
-                var cy = primaryDetection.center.Y - cropHeight / 2f;
+                var cx = primaryDetection.Center.X - cropWidth  / 2f;
+                var cy = primaryDetection.Center.Y - cropHeight / 2f;
 
                 var r = new Rect(cx, cy, cropWidth, cropHeight);
 
                 crop = ClampCrop(r, w, h);
             }
 
-            var boxes = detections.Select(x => x.box).ToList();
-            Preview = new PreviewData(frame, boxes, crop, Job.GravitateTo ?? new (0.5f, 0.5f));
+            var boxes = detections.Select(x => x.Box).ToList();
+            Preview = new PreviewData(frame, boxes, crop, Job.GravitateTo, pos, Job.Rotate);
         }
         catch (Exception ex)
         {

Splitter-UI/Views/InspectorPane.axaml

@@ -86,6 +86,25 @@ x:DataType="vm:InspectorPaneViewModel">
                               Width="160"/>
                 </StackPanel>
 
+                <!-- ScoreThreshold -->
+                <StackPanel Orientation="Horizontal" Spacing="8">
+                    <TextBlock Text="Score Threshold" Width="120"/>
+
+                    <StackPanel Orientation="Vertical" Spacing="4" Width="260">
+                        <Slider Minimum="0"
+                                Maximum="1"
+                                SmallChange="0.01"
+                                LargeChange="0.1"
+                                TickFrequency="0.05"
+                                IsSnapToTickEnabled="False"
+                                Value="{Binding Selected.ScoreThreshold, Mode=TwoWay}"/>
+
+                        <TextBlock Text="{Binding Selected.ScoreThreshold, StringFormat='0.00'}"
+                                   FontSize="10"
+                                   HorizontalAlignment="Right"/>
+                    </StackPanel>
+                </StackPanel>
+
                 <!-- DetectAbove -->
                 <StackPanel Orientation="Horizontal" Spacing="8">
                     <TextBlock Text="Detect Above" Width="120"/>

Splitter-UI/Views/MainWindow.axaml

@@ -8,7 +8,7 @@
     x:DataType="vm:MainViewModel"
     x:Name="Root"
     Width="1800"
-    Height="790"
+    Height="830"
     Title="Splitter UI"
     Icon="avares://Splitter-UI/Assets/splitter.png">
 

splitter-cli/CommandLine.cs

@@ -98,6 +98,14 @@ public sealed class CommandLine
                 else
                     Master.DetectAbove = 0.7f;
             }
+            else if (arg.StartsWith("--score-threshold="))
+            {
+                var val = arg.Substring("--score-threshold=".Length);
+                if (float.TryParse(val, NumberStyles.Float, CultureInfo.InvariantCulture, out var scoreThreshold) && scoreThreshold >= 0.0f && scoreThreshold <= 1.0f)
+                    Master.ScoreThreshold = scoreThreshold;
+                else
+                    Master.ScoreThreshold = 0.25f;
+            }
             else if (arg == "--crop")
             {
                 Master.Crop = ParseCrop("");
@@ -224,22 +232,22 @@ public sealed class CommandLine
         return key.Length > 0;
     }
 
-    private static Point2f? ParseGravitate(string value)
+    private static Point2f ParseGravitate(string value)
     {
         // Expected format: "<x>:<y>"
         var parts = value.Split(':');
         if (parts.Length != 2)
-            return null;
+            return new Point2f(0.5f, 0.5f);
 
         if (!float.TryParse(parts[0], NumberStyles.Float, CultureInfo.InvariantCulture, out var x))
-            return null;
+            return new Point2f(0.5f, 0.5f);
 
         if (!float.TryParse(parts[1], NumberStyles.Float, CultureInfo.InvariantCulture, out var y))
-            return null;
+            return new Point2f(0.5f, 0.5f);
 
         // Normalized range check (0.0–1.0)
         if (x < 0f || x > 1f || y < 0f || y > 1f)
-            return null;
+            return new Point2f(0.5f, 0.5f);
 
         return new Point2f(x, y);
     }

splitter-cli/JobProcessor.cs

@@ -64,11 +64,13 @@ public class JobProcessor(ILogger logger) : LoggingBase(logger, 0), IJobProcesso
                 IObjectDetector detector = job.Detect switch
                 {
                     "face" => new UltraFaceDetector(_logger),
-                    "body" => new YoloOnnxObjectDetector(_logger),
+                    "body" => new YoloV10ObjectDetector(_logger),
                     "none" => new DummyDetector(),
                     _      => throw new InvalidOperationException($"Unknown detector: {job.Detect}")
                 };
-                return new TrackingSplitter(i, detector, job, _logger);
+                var osnet = new OSNetEmbeddingExtractor();
+                var tracker = new ObjectTracker(detector, osnet);
+                return new TrackingSplitter(i, tracker, job, _logger);
             };
         }
         else

splitter-cli/SingleJob.cs

@@ -29,12 +29,7 @@ public class SingleJob
     /// such as left-center (0.2, 0.5) or top-right (0.8, 0.2). This can be useful for 
     /// videos where the subject tends to be off-center or for creative framing choices.
     /// </summary>
-    public Point2f?                   GravitateTo            { get; set; }
-    /// <summary>
-    /// Face or human detectors should only report detections if their upper bound starts below this threshold. 
-    /// This is a value between 0.0 and 1.0 mapped to 0..Height.
-    /// </summary>
-    public float                      DetectAbove            { get; set; } = 0.3f;
+    public Point2f                    GravitateTo            { get; set; } = new Point2f(0.5f, 0.5f);
     /// <summary>
     /// Destination file mask.
     /// </summary>
@@ -50,6 +45,15 @@ public class SingleJob
     /// </summary>
     public string?                    Detect                 { get; set; }
     /// <summary>
+    /// Detection confidence threshold. This is a value between 0.0 and 1.0 that sets the minimum confidence
+    /// </summary>
+    public float                      ScoreThreshold         { get; set; } = 0.25f;
+    /// <summary>
+    /// Face or human detectors should only report detections if their upper bound starts below this threshold. 
+    /// This is a value between 0.0 and 1.0 mapped to 0..Height.
+    /// </summary>
+    public float                      DetectAbove            { get; set; } = 0.7f;
+    /// <summary>
     /// Set starget segments length explicitly. By default, the splitter calculates segment 
     /// lengths to be equal and not exceed 58 seconds.
     /// </summary>

splitter-cli/TrackingSplitter.cs

@@ -4,24 +4,18 @@ using System.Runtime.InteropServices;
 
 namespace splitter;
 
-public class TrackingSplitter : LoggingBase, ISegmentProcessor, IDisposable
+public class TrackingSplitter : LoggingBase, ISegmentProcessor
 {
-    private readonly IObjectDetector _detector;
+    private readonly IObjectTracker          _tracker;
 
     public TrackingSplitter(
         int  progressLine,
-        IObjectDetector detector,
+        IObjectTracker tracker,
         SingleJob cmd,
         ILogger logger)
         : base(logger, progressLine)
     {
-        _detector     = detector;
-    }
-
-    public void Dispose()
-    {
-        if (_detector is IDisposable d)
-            d.Dispose();
+        _tracker = tracker;
     }
 
     public async Task ProcessSegment(SingleTask job, CancellationToken token)
@@ -103,12 +97,12 @@ public class TrackingSplitter : LoggingBase, ISegmentProcessor, IDisposable
 
         var kalman = new KalmanTracker();
         var camera = new CameraController(
-        videoWidth,
-        videoHeight,
-        job.Job.Crop.Value.width,
-        job.Job.Crop.Value.height,
-        kalman,
-        job.Job);
+            videoWidth,
+            videoHeight,
+            job.Job.Crop.Value.width,
+            job.Job.Crop.Value.height,
+            kalman,
+            job.Job);
 
         try
         {
@@ -130,12 +124,7 @@ public class TrackingSplitter : LoggingBase, ISegmentProcessor, IDisposable
 
                 Marshal.Copy(inBuffer, 0, frameMat.Data, inBytes);
 
-                var objects = _detector.DetectAll(job, frameMat);
-
-                // Ignore detections starting in the lower 1/2 of the frame
-                objects = objects.Where(o => o.center.Y <= frameMat.Height * job.Job.DetectAbove).ToList();
-
-                var primary = SelectTrackedObject(objects, kalman.LastMeasurement);
+                var (objects, primary) = _tracker.SelectTrackedObject(job, frameMat, kalman.LastMeasurement);
 
                 camera.Update(primary);
                 var roi = camera.Roi;
@@ -389,7 +378,7 @@ public class TrackingSplitter : LoggingBase, ISegmentProcessor, IDisposable
 
     private void DrawDebug(
         Mat frame,
-        System.Collections.Generic.List<(Rect box, Point2f center)> objects,
+        List<DetectedPerson> objects,
         CameraController camera,
         KalmanTracker kalman)
     {
@@ -418,52 +407,4 @@ public class TrackingSplitter : LoggingBase, ISegmentProcessor, IDisposable
             HersheyFonts.HersheySimplex, 0.6, color, 2);
     }
 
-    private (Rect box, Point2f center)? SelectTrackedObject(
-        List<(Rect box, Point2f center)> foundObjects,
-        Point2f? previousCenter)
-    {
-        if (foundObjects == null || foundObjects.Count == 0)
-            return null;
-
-        if (!previousCenter.HasValue)
-        {
-            var bestIndex = 0;
-            var bestArea  = float.MinValue;
-
-            for (var i = 0; i < foundObjects.Count; i++)
-            {
-                var f    = foundObjects[i];
-                var area = f.box.Width * f.box.Height;
-                if (area > bestArea)
-                {
-                    bestArea = area;
-                    bestIndex = i;
-                }
-            }
-
-            return foundObjects[bestIndex];
-        }
-        else
-        {
-            var prev      = previousCenter.Value;
-            var bestIndex = 0;
-            var bestDist2 = float.MaxValue;
-
-            for (var i = 0; i < foundObjects.Count; i++)
-            {
-                var f  = foundObjects[i];
-                var dx = f.center.X - prev.X;
-                var dy = f.center.Y - prev.Y;
-                var d2 = dx * dx + dy * dy;
-
-                if (d2 < bestDist2)
-                {
-                    bestDist2 = d2;
-                    bestIndex = i;
-                }
-            }
-
-            return foundObjects[bestIndex];
-        }
-    }
 }

splitter-cli/algo/CameraController.cs

@@ -58,7 +58,7 @@ public sealed class CameraController
         _kalman.Reset(_cameraCenter);
     }
 
-    private Point2f DefaultCenter => _cmd.GravitateTo ?? new Point2f(_videoWidth / 2f, _videoHeight / 2f);
+    private Point2f DefaultCenter => _cmd.GravitateTo;
 
     public int LostFrames         => _lostFrames;
     public Point2f CameraCenter   => _cameraCenter;
@@ -68,15 +68,15 @@ public sealed class CameraController
     public Point2f? ObjectCenter  => _objectCenter;
     public Rect Roi => _roi;
 
-    public void Update((Rect box, Point2f center)? primary)
+    public void Update(DetectedPerson? primary)
     {
         Rect?    objectBox    = null;
         Point2f? objectCenter = null;
 
         if (primary.HasValue)
         {
-            objectCenter = primary.Value.center;
-            objectBox = primary.Value.box;
+            objectCenter = primary.Value.Center;
+            objectBox = primary.Value.Box;
         }
 
         // ---------------------------------------------------------

splitter-cli/algo/DetectedPerson.cs

@@ -0,0 +1,8 @@
+namespace splitter.algo;
+
+public struct DetectedPerson
+{
+    public ulong   Id;
+    public Rect    Box;
+    public Point2f Center;
+}

splitter-cli/algo/DummyDetector.cs

@@ -2,19 +2,19 @@
 
 public sealed class DummyDetector : IObjectDetector
 {
-    public List<(Rect box, Point2f center)> DetectAll(SingleTask job, Mat frameCont)
+    public List<DetectedPerson> DetectAll(SingleTask job, Mat frameCont)
     {
         var h   = job.Info.Height;
         var w   = job.Info.Width;
 
-        var c = job.Job.GravitateTo ?? new Point2f(0.5f, 0.5f);
+        var c = job.Job.GravitateTo;
         var x = (int)(c.X * w);
         var y = (int)(c.Y * h);
 
         var center = new Point2f(x, y);
         var rect = new Rect(x - 1, y - 1, 2, 2);
 
-        return [(rect, center)];
+        return [new DetectedPerson { Box = rect, Center = center }];
     }
 
     public void Dispose() {}

splitter-cli/algo/IEmbeddingExtractor.cs

@@ -0,0 +1,6 @@
+namespace splitter.algo;
+
+public interface IEmbeddingExtractor : IDisposable
+{
+    float[] Extract(Mat frame, Rect box);
+}

splitter-cli/algo/IObjectDetector.cs

@@ -2,5 +2,5 @@
 
 public interface IObjectDetector : IDisposable
 {
-    List<(Rect box, Point2f center)> DetectAll(SingleTask job, Mat frameCont);
+    List<DetectedPerson> DetectAll(SingleTask job, Mat frameCont);
 }

splitter-cli/algo/IObjectTracker.cs

@@ -0,0 +1,6 @@
+namespace splitter.algo;
+
+public interface IObjectTracker
+{
+    (List<DetectedPerson>, DetectedPerson?) SelectTrackedObject(SingleTask job, Mat frameMat, Point2f? lastMeasurement);
+}

splitter-cli/algo/OSNetEmbeddingExtractor.cs

@@ -0,0 +1,127 @@
+using System.Runtime.CompilerServices;
+using Microsoft.ML.OnnxRuntime;
+using Microsoft.ML.OnnxRuntime.Tensors;
+
+namespace splitter.algo;
+
+public sealed class OSNetEmbeddingExtractor : IDisposable, IEmbeddingExtractor
+{
+    private readonly InferenceSession _session;
+    private readonly string _inputName;
+    private readonly string _outputName;
+
+    private const int _batchSize    = 16;
+    private const int _inputHeight  = 256;
+    private const int _inputWidth   = 128;
+    private const int _channels     = 3;
+
+    private readonly float[] _inputBuffer;
+    private readonly DenseTensor<float> _inputTensor;
+    private readonly List<NamedOnnxValue> _inputs = new(1);
+
+    private readonly float[] _embedding;
+
+    private readonly Mat _resizeMat = new();
+    private readonly Mat _rgbMat    = new();
+
+    private readonly float _inv255 = 1f / 255f;
+
+    public OSNetEmbeddingExtractor()
+    {
+        var opt = new SessionOptions();
+        opt.AppendExecutionProvider_DML();
+
+        var modelPath = Path.Combine(AppDomain.CurrentDomain.BaseDirectory, "models", "osnet_x0_25_msmt17.onnx");
+        _session = new InferenceSession(modelPath, opt);
+
+        _inputName = _session.InputMetadata.Keys.First();
+        _outputName = _session.OutputMetadata.Keys.First();
+
+        int inputSize = _batchSize * _channels * _inputHeight * _inputWidth;
+        _inputBuffer = new float[inputSize];
+
+        _inputTensor = new DenseTensor<float>(
+            _inputBuffer,
+            new[] { _batchSize, _channels, _inputHeight, _inputWidth }
+        );
+
+        _inputs.Add(NamedOnnxValue.CreateFromTensor(_inputName, _inputTensor));
+
+        int outDim = _session.OutputMetadata[_outputName].Dimensions[1];
+        _embedding = new float[outDim];
+    }
+
+    public float[] Extract(Mat frame, Rect box)
+    {
+        // Clear all batches
+        Array.Clear(_inputBuffer, 0, _inputBuffer.Length);
+
+        // Extract ROI
+        var roi = new Mat(frame, box);
+
+        Cv2.Resize(roi, _resizeMat, new Size(_inputWidth, _inputHeight));
+        Cv2.CvtColor(_resizeMat, _rgbMat, ColorConversionCodes.BGR2RGB);
+
+        FillBatch0(_rgbMat);
+
+        using var results = _session.Run(_inputs);
+
+        var output = results.First(v => v.Name == _outputName).AsTensor<float>();
+
+        // Read embedding from batch 0
+        for (int i = 0; i < _embedding.Length; i++)
+            _embedding[i] = output[0, i];
+
+        NormalizeL2(_embedding);
+
+        return _embedding;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private void FillBatch0(Mat rgb)
+    {
+        int plane = _inputHeight * _inputWidth;
+
+        unsafe
+        {
+            for (int y = 0; y < _inputHeight; y++)
+            {
+                var rowPtr  = (byte*)rgb.Ptr(y).ToPointer();
+                var rowSpan = new Span<byte>(rowPtr, _inputWidth * 3);
+
+                int src = 0;
+
+                for (int x = 0; x < _inputWidth; x++)
+                {
+                    int off = y * _inputWidth + x;
+
+                    _inputBuffer[off] = rowSpan[src + 0] * _inv255; // R
+                    _inputBuffer[plane + off] = rowSpan[src + 1] * _inv255; // G
+                    _inputBuffer[2 * plane + off] = rowSpan[src + 2] * _inv255; // B
+
+                    src += 3;
+                }
+            }
+        }
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static void NormalizeL2(float[] v)
+    {
+        float sum = 0f;
+        for (int i = 0; i < v.Length; i++)
+            sum += v[i] * v[i];
+
+        float inv = 1f / MathF.Sqrt(sum);
+
+        for (int i = 0; i < v.Length; i++)
+            v[i] *= inv;
+    }
+
+    public void Dispose()
+    {
+        _session?.Dispose();
+        _resizeMat?.Dispose();
+        _rgbMat?.Dispose();
+    }
+}

splitter-cli/algo/ObjectTracker.cs

@@ -0,0 +1,98 @@
+namespace splitter.algo;
+
+public class ObjectTracker(IObjectDetector _detector, IEmbeddingExtractor _embeddingExtractor) : IObjectTracker
+{
+    public (List<DetectedPerson> /*objects*/, DetectedPerson? /*primary*/) SelectTrackedObject(SingleTask job, Mat frameMat, Point2f? lastMeasurement)
+    {
+        var objects = _detector.DetectAll(job, frameMat) ?? [];
+
+        // Ignore detections starting in the lower 1/2 of the frame
+        objects = objects.Where(o => o.Center.Y <= frameMat.Height * job.Job.DetectAbove).ToList();
+
+        // attach embeddings to all persons
+        for (int i = 0; i < objects.Count; i++)
+        {
+            var p = objects[i];   // copy struct
+
+            var rect = p.Box;
+
+            rect.X      = Math.Clamp(rect.X, 0, frameMat.Width - 1);
+            rect.Y      = Math.Clamp(rect.Y, 0, frameMat.Height - 1);
+            rect.Width  = Math.Clamp(rect.Width, 1, frameMat.Width - rect.X);
+            rect.Height = Math.Clamp(rect.Height, 1, frameMat.Height - rect.Y);
+
+            var embedding = _embeddingExtractor.Extract(frameMat, rect);
+            p.Id = HashEmbedding(embedding); // assign ID based on embedding hash
+
+            objects[i] = p;       // write back
+        }
+
+        var primary = SelectPrimaryObject(objects, lastMeasurement);
+        return (objects, primary);
+    }
+
+    private static ulong HashEmbedding(float[] emb)
+    {
+        unchecked
+        {
+            ulong hash = 146527;
+            for (int i = 0; i < emb.Length; i++)
+            {
+                // convert float to int bits
+                uint bits = (uint)BitConverter.SingleToInt32Bits(emb[i]);
+                hash = (hash * 16777619) ^ bits;
+            }
+            return hash;
+        }
+    }
+
+    private DetectedPerson? SelectPrimaryObject(
+        List<DetectedPerson> foundObjects,
+        Point2f? previousCenter)
+    {
+        if (foundObjects == null || foundObjects.Count == 0)
+            return null;
+
+        if (!previousCenter.HasValue)
+        {
+            var bestIndex = 0;
+            var bestArea  = float.MinValue;
+
+            for (var i = 0; i < foundObjects.Count; i++)
+            {
+                var f    = foundObjects[i];
+                var area = f.Box.Width * f.Box.Height;
+                if (area > bestArea)
+                {
+                    bestArea = area;
+                    bestIndex = i;
+                }
+            }
+
+            return foundObjects[bestIndex];
+        }
+        else
+        {
+            var prev      = previousCenter.Value;
+            var bestIndex = 0;
+            var bestDist2 = float.MaxValue;
+
+            for (var i = 0; i < foundObjects.Count; i++)
+            {
+                var f  = foundObjects[i];
+                var dx = f.Center.X - prev.X;
+                var dy = f.Center.Y - prev.Y;
+                var d2 = dx * dx + dy * dy;
+
+                if (d2 < bestDist2)
+                {
+                    bestDist2 = d2;
+                    bestIndex = i;
+                }
+            }
+
+            return foundObjects[bestIndex];
+        }
+    }
+
+}

splitter-cli/algo/UltraFaceDetector.cs

@@ -23,14 +23,14 @@ public sealed class UltraFaceDetector: LoggingBase, IDisposable, IObjectDetector
         _ultraFace = UltraFace.Create(param);
     }
 
-    public List<(Rect box, Point2f center)> DetectAll(SingleTask job, Mat frameCont)
+    public List<DetectedPerson> DetectAll(SingleTask job, Mat frameCont)
     {
         // Convert to byte[] for UltraFace
         var bytesFull = frameCont.Rows * frameCont.Cols * frameCont.ElemSize();
         var bgr = new byte[bytesFull];
         Marshal.Copy(frameCont.Data, bgr, 0, bytesFull);
 
-        var results = new List<(Rect box, Point2f center)>();
+        var results = new List<DetectedPerson>();
 
         if (bgr == null || bgr.Length == 0)
             return results;
@@ -69,7 +69,7 @@ public sealed class UltraFaceDetector: LoggingBase, IDisposable, IObjectDetector
                     rect.X + rect.Width / 2f,
                     rect.Y + rect.Height / 2f);
 
-                    results.Add((rect, center));
+                    results.Add(new DetectedPerson{ Box = rect, Center = center });
                 }
             }
         }

splitter-cli/algo/YoloV10ObjectDetector.cs

@@ -0,0 +1,278 @@
+using System.Runtime.CompilerServices;
+using Microsoft.ML.OnnxRuntime;
+using Microsoft.ML.OnnxRuntime.Tensors;
+
+namespace splitter.algo;
+
+public sealed class YoloV10ObjectDetector : LoggingBase, IObjectDetector, IDisposable
+{
+    private readonly InferenceSession _session;
+    private readonly string _inputName;
+    private readonly string _outputName;
+
+    private const int   _inputWidth       = 640;
+    private const int   _inputHeight      = 640;
+    private const float _scoreThreshold   = 0.35f;
+    private const float _nmsThreshold     = 0.45f;
+    private const int   _personClassIndex = 0;
+
+    private readonly Mat _resizeMat = new();
+    private readonly Mat _rgbMat    = new();
+
+    private readonly float[] _inputBuffer;
+    private readonly DenseTensor<float> _inputTensor;
+
+    private readonly List<NamedOnnxValue> _inputs = new(1);
+
+    private readonly List<Detection> _detections = new(256);
+    private readonly List<Detection> _nmsBuffer  = new(256);
+
+    private readonly List<DetectedPerson> _results = new(64);
+
+    private readonly float _inv255 = 1f / 255f;
+
+    private readonly struct Detection
+    {
+        public readonly float X;
+        public readonly float Y;
+        public readonly float Width;
+        public readonly float Height;
+        public readonly float Score;
+
+        public Detection(float x, float y, float w, float h, float score)
+        {
+            X = x;
+            Y = y;
+            Width = w;
+            Height = h;
+            Score = score;
+        }
+    }
+
+    public YoloV10ObjectDetector(ILogger logger) : base(logger, -1)
+    {
+        var options = new SessionOptions();
+        options.AppendExecutionProvider_DML();
+
+        var basePath  = AppDomain.CurrentDomain.BaseDirectory;
+        var modelPath = Path.Combine(basePath, "models", "yolov10m.onnx");
+
+        _session = new InferenceSession(modelPath, options);
+
+        _inputName = _session.InputMetadata.Keys.First();
+        _outputName = _session.OutputMetadata.Keys.First();
+
+        _inputBuffer = new float[1 * 3 * _inputHeight * _inputWidth];
+        _inputTensor = new DenseTensor<float>(_inputBuffer, new[] { 1, 3, _inputHeight, _inputWidth });
+
+        _inputs.Add(NamedOnnxValue.CreateFromTensor(_inputName, _inputTensor));
+    }
+
+    public List<DetectedPerson> DetectAll(SingleTask job, Mat frameCont)
+    {
+        if (frameCont.Empty())
+        {
+            _results.Clear();
+            return _results;
+        }
+
+        Cv2.Resize(frameCont, _resizeMat, new Size(_inputWidth, _inputHeight));
+        Cv2.CvtColor(_resizeMat, _rgbMat, ColorConversionCodes.BGR2RGB);
+
+        FillInputTensor(_rgbMat);
+
+        using var results = _session.Run(_inputs);
+
+        Tensor<float>? output = null;
+        foreach (var r in results)
+        {
+            if (r.Name == _outputName)
+            {
+                output = r.AsTensor<float>();
+                break;
+            }
+        }
+
+        if (output is null)
+        {
+            _results.Clear();
+            return _results;
+        }
+
+        ParseYoloV10(
+            output,
+            frameCont.Width,
+            frameCont.Height,
+            job.Job.ScoreThreshold,
+            _personClassIndex,
+            _detections);
+
+        var final = ApplyNms(_detections, _nmsThreshold, _nmsBuffer);
+
+        _results.Clear();
+        for (var i = 0; i < final.Count; i++)
+        {
+            var d = final[i];
+
+            var x = (int)d.X;
+            var y = (int)d.Y;
+            var w = (int)d.Width;
+            var h = (int)d.Height;
+
+            x = Math.Clamp(x, 0, frameCont.Width - 1);
+            y = Math.Clamp(y, 0, frameCont.Height - 1);
+            w = Math.Clamp(w, 1, frameCont.Width - x);
+            h = Math.Clamp(h, 1, frameCont.Height - y);
+
+            var rect   = new Rect(x, y, w, h);
+            var center = new Point2f(x + w / 2f, y + h / 2f);
+
+            _results.Add(new DetectedPerson{ Box = rect, Center = center });
+        }
+
+        return _results;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private void FillInputTensor(Mat rgb)
+    {
+        var height = _inputHeight;
+        var width  = _inputWidth;
+
+        var planeSize = height * width;
+
+        Span<float> dst = _inputBuffer.AsSpan();
+
+        unsafe
+        {
+            for (var y = 0; y < height; y++)
+            {
+                var rowPtr  = (byte*)rgb.Ptr(y).ToPointer();
+                var rowSpan = new Span<byte>(rowPtr, width * 3);
+
+                var srcIndex = 0;
+
+                for (var x = 0; x < width; x++)
+                {
+                    var r = rowSpan[srcIndex + 0];
+                    var g = rowSpan[srcIndex + 1];
+                    var b = rowSpan[srcIndex + 2];
+
+                    var offset = y * width + x;
+
+                    dst[offset] = r * _inv255;
+                    dst[planeSize + offset] = g * _inv255;
+                    dst[2 * planeSize + offset] = b * _inv255;
+
+                    srcIndex += 3;
+                }
+            }
+        }
+    }
+
+    // YOLOv10 parser: [1, 300, 6] => x1, y1, x2, y2, score, class_id
+    private static void ParseYoloV10(
+        Tensor<float> output,
+        int originalWidth,
+        int originalHeight,
+        float scoreThreshold,
+        int classIndex,
+        List<Detection> detections)
+    {
+        detections.Clear();
+
+        // dims: [1, 300, 6]
+        var count = output.Dimensions[1];
+
+        var xScale = (float)originalWidth  / 640f;
+        var yScale = (float)originalHeight / 640f;
+
+        for (var i = 0; i < count; i++)
+        {
+            var x1    = output[0, i, 0];
+            var y1    = output[0, i, 1];
+            var x2    = output[0, i, 2];
+            var y2    = output[0, i, 3];
+            var score = output[0, i, 4];
+            var cls   = (int)output[0, i, 5];
+
+            if (cls != classIndex)
+                continue;
+
+            if (score < scoreThreshold)
+                continue;
+
+            var left   = x1 * xScale;
+            var top    = y1 * yScale;
+            var width  = (x2 - x1) * xScale;
+            var height = (y2 - y1) * yScale;
+
+            detections.Add(new Detection(left, top, width, height, score));
+        }
+    }
+
+    private static List<Detection> ApplyNms(
+        List<Detection> detections,
+        float nmsThreshold,
+        List<Detection> nmsBuffer)
+    {
+        nmsBuffer.Clear();
+
+        if (detections.Count == 0)
+            return nmsBuffer;
+
+        detections.Sort(static (a, b) => b.Score.CompareTo(a.Score));
+
+        for (var i = 0; i < detections.Count; i++)
+        {
+            var candidate = detections[i];
+            var keep = true;
+
+            for (var j = 0; j < nmsBuffer.Count; j++)
+            {
+                if (IoU(candidate, nmsBuffer[j]) >= nmsThreshold)
+                {
+                    keep = false;
+                    break;
+                }
+            }
+
+            if (keep)
+                nmsBuffer.Add(candidate);
+        }
+
+        return nmsBuffer;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static float IoU(in Detection a, in Detection b)
+    {
+        var x1 = MathF.Max(a.X, b.X);
+        var y1 = MathF.Max(a.Y, b.Y);
+        var x2 = MathF.Min(a.X + a.Width,  b.X + b.Width);
+        var y2 = MathF.Min(a.Y + a.Height, b.Y + b.Height);
+
+        var interW = x2 - x1;
+        if (interW <= 0f) return 0f;
+
+        var interH = y2 - y1;
+        if (interH <= 0f) return 0f;
+
+        var interArea = interW * interH;
+
+        var areaA = a.Width * a.Height;
+        var areaB = b.Width * b.Height;
+
+        var union = areaA + areaB - interArea;
+        if (union <= 0f) return 0f;
+
+        return interArea / union;
+    }
+
+    public void Dispose()
+    {
+        _session?.Dispose();
+        _resizeMat?.Dispose();
+        _rgbMat?.Dispose();
+    }
+}

splitter-cli/algo/YoloV8ObjectDetector.cs

@@ -4,7 +4,7 @@ using Microsoft.ML.OnnxRuntime.Tensors;
 
 namespace splitter.algo;
 
-public sealed class YoloOnnxObjectDetector : LoggingBase, IObjectDetector, IDisposable
+public sealed class YoloV8ObjectDetector : LoggingBase, IObjectDetector, IDisposable
 {
     private readonly InferenceSession _session;
     private readonly string _inputName;
@@ -32,7 +32,7 @@ public sealed class YoloOnnxObjectDetector : LoggingBase, IObjectDetector, IDisp
     private readonly List<Detection> _nmsBuffer  = new(256);
 
     // Reusable result list
-    private readonly List<(Rect box, Point2f center)> _results = new(64);
+    private readonly List<DetectedPerson> _results = new(64);
 
     private readonly float _inv255 = 1f / 255f;
 
@@ -54,7 +54,7 @@ public sealed class YoloOnnxObjectDetector : LoggingBase, IObjectDetector, IDisp
         }
     }
 
-    public YoloOnnxObjectDetector(ILogger logger) : base(logger, -1)
+    public YoloV8ObjectDetector(ILogger logger) : base(logger, -1)
     {
         var options = new SessionOptions();
         options.AppendExecutionProvider_DML();
@@ -78,7 +78,7 @@ public sealed class YoloOnnxObjectDetector : LoggingBase, IObjectDetector, IDisp
         _inputs.Add(NamedOnnxValue.CreateFromTensor(_inputName, _inputTensor));
     }
 
-    public List<(Rect box, Point2f center)> DetectAll(SingleTask job, Mat frameCont)
+    public List<DetectedPerson> DetectAll(SingleTask job, Mat frameCont)
     {
         if (frameCont.Empty())
         {
@@ -142,7 +142,7 @@ public sealed class YoloOnnxObjectDetector : LoggingBase, IObjectDetector, IDisp
             var rect   = new Rect(x, y, w, h);
             var center = new Point2f(x + w / 2f, y + h / 2f);
 
-            _results.Add((rect, center));
+            _results.Add(new DetectedPerson{ Box = rect, Center = center });
         }
 
         return _results;