LMSupply.Detector 0.63.0

LMSupply.Detector

Local object detection for .NET with automatic model downloading.

Features

• Zero-config: Models download automatically from HuggingFace
• GPU Acceleration: CUDA, DirectML (Windows), CoreML (macOS)
• Permissively licensed: RT-DETR (Apache-2.0) and YuNet (MIT), redistributable in a closed-source commercial product. No alias resolves to an AGPL-3.0 YOLO checkpoint.
• Per-model vocabulary: COCO-80 (people, vehicles, animals, objects) by default, and a model that was trained on something else carries its own labels — DetectorModelInfo.ClassLabels, with NumClasses derived from it so the two cannot disagree. Post-processing labels from the loaded model, not from COCO by assumption; an id outside the vocabulary reads as unknown rather than borrowing a COCO name.

Quick Start

using LMSupply.Detector;

// Load the default model
await using var detector = await LocalDetector.LoadAsync("default");

// Detect objects
var results = await detector.DetectAsync("photo.jpg");

foreach (var detection in results)
{
    Console.WriteLine($"{detection.Label}: {detection.Confidence:P1}");
    Console.WriteLine($"  Box: [{detection.Box.X1:F0}, {detection.Box.Y1:F0}]");
}

Available Models

Faces

face resolves to OpenCV's YuNet. COCO has no face class and person is not a substitute for redaction work, so this is a separate model with its own single-label vocabulary and its own decoder.

await using var detector = await LocalDetector.LoadAsync("face");

foreach (var face in await detector.DetectAsync("photo.jpg"))
{
    // face.Label is "face"; face.Keypoints holds the two eyes, the nose tip and the two mouth corners,
    // each carrying the detection's own score - YuNet publishes no per-landmark confidence.
    Console.WriteLine($"{face.Box.X1:F0},{face.Box.Y1:F0} {face.Box.Width:F0}x{face.Box.Height:F0}");
}

Feed it ordinary images: it wants BGR bytes rather than the scaled RGB the RT-DETR aliases take, and that conversion happens inside the library.

Measured cost (1280x1177 JPEG, 4-core CPU): about 19 ms per frame end to end, of which roughly half is JPEG decoding - the model itself runs in about 2.4 ms. Detection is therefore comfortably inside a 30 fps budget, and JPEG decoding is the thing to avoid paying twice for if frames arrive already decoded. DirectML rejects one of this model's operators, so it runs on CPU even on a machine where the RT-DETR aliases get a GPU; the provider fallback handles this without configuration.

The defaults suit it: ConfidenceThreshold 0.25 and IouThreshold 0.45. A stricter IouThreshold of 0.3 matches the reference implementation and merges one more duplicate in a dense crowd; the difference measured on a street scene was one box out of eight.

Licence plates

plate resolves to OpenCV's licence-plate YuNet. It shares a name with the face model and little else - 320x240 input, prior boxes rather than an anchor-free grid, and a quadrilateral rather than an upright box, because a plate photographed from an angle is not axis-aligned.

await using var detector = await LocalDetector.LoadAsync("plate");

foreach (var plate in await detector.DetectAsync("photo.jpg"))
{
    // plate.Box is the upright hull - what you want if you are blurring the region.
    // plate.Keypoints holds the four corners, clockwise from top-left - what you want if you are
    // rectifying the plate to read it, or blurring a tighter polygon.
}

Measured cost (960x631 JPEG, DirectML on an integrated GPU): about 7 ms per frame end to end; on CPU the model alone is 5.3 ms. Unlike face, this one does run on DirectML.

The library defaults (ConfidenceThreshold 0.25, IouThreshold 0.45) are usable: measured plates scored 0.63-0.99 while a cat photograph and a crowded street scene both produced nothing at all, the highest score anywhere in them being 0.15. The reference implementation uses a stricter 0.8 with an IoU of 0.3; raise the threshold if false positives cost you more than misses.

GPU Acceleration

# NVIDIA GPU
dotnet add package Microsoft.ML.OnnxRuntime.Gpu

# Windows (AMD/Intel/NVIDIA)
dotnet add package Microsoft.ML.OnnxRuntime.DirectML