Dartface #

Dartface detects mostly frontal faces in still photographs with a pure Dart detection engine for Flutter. It returns source-pixel bounding boxes, six estimated landmarks, confidence scores, and in-plane rotation without TensorFlow Lite, FFI or platform channels.

The default detector evaluates a trained local-binary-pattern cascade and groups neighboring positive windows. Encoded files are decoded by the Focale imcodec package. No image or face data leaves the device.

Features #

• Pure Dart multi-scale LBP cascade evaluation.
• JPEG, PNG, WebP, GIF, BMP, TIFF, TGA, QOI, and JPEG XL input through Imcodec.
• Direct imcodec.Image, grayscale, RGB, BGR, RGBA, BGRA, ARGB, ABGR, and planar YUV420 input.
• Background-isolate decoding and detection on native Dart and Flutter platforms.
• Inline execution on the web, where isolates are unavailable.
• Immutable results and owned image buffers.
• Configurable accuracy, face-size bounds, box and landmark confidence thresholds, result limit, and decoded-pixel safety limit.
• Six lightweight landmark estimates: both eyes, nose tip, and three mouth points.
• Optional accurate-mode recovery for faces tilted about 15 degrees in either direction.

Installation #

Add Dartface to a Flutter package:

flutter pub add dartface

Detect an encoded image #

import 'dart:typed_data';

import 'package:dartface/dartface.dart';

Future<List<DetectedFace>> findFaces(Uint8List jpegBytes) async {
  final FaceDetector detector = await FaceDetector.create(
    options: const FaceDetectorOptions(
      accuracy: FaceDetectionAccuracy.balanced,
      minimumFaceSize: 0.15,
      confidenceThreshold: 0.58,
    ),
  );
  try {
    return await detector.detectFacesFromBytes(jpegBytes);
  } finally {
    await detector.dispose();
  }
}

Each result uses the source image's pixel coordinate system:

for (final DetectedFace face in faces) {
  print(
    '${face.boundingBox}: box=${face.confidence}, '
    'landmarks=${face.landmarkConfidence}',
  );
  print(face.landmark(FaceLandmarkType.leftEye)?.position);
}

Neither confidence value is a calibrated probability. confidence ranks support for the face box, while landmarkConfidence is the mean confidence of the six sparse feature positions. The default options require both to be usable. In accurate mode, the held-out portion of the small reference corpus has a mean landmark error of 7.3% of the eye spacing; mean roll error across all labelled accurate-mode faces is 1.5 degrees.

Detect decoded pixels #

Use FaceImage when pixels already come from an editor or another imaging pipeline:

final FaceImage frame = FaceImage.fromBytes(
  width: width,
  height: height,
  bytes: bgraBytes,
  pixelFormat: FacePixelFormat.bgra8888,
  rowStride: bytesPerRow,
);

final List<DetectedFace> faces = await detector.detectFacesAsync(frame);

Planar YUV420 input is also accepted for interoperability, although Dartface is not tuned for continuous video:

final FaceImage frame = FaceImage.fromYuv420(
  width: width,
  height: height,
  luminancePlane: yPlane,
  chromaBluePlane: uPlane,
  chromaRedPlane: vPlane,
  luminanceRowStride: yRowStride,
  chromaRowStride: uvRowStride,
  chromaPixelStride: uvPixelStride,
);

If the caller already has an imcodec.Image, detectFacesFromImage avoids an encode/decode cycle.

FaceImage.rgbaBytes returns a defensive copy. When pixels are only read, such as when handing a frame to another library, FaceImage.rgbaView returns an unmodifiable view over the same buffer instead.

Options #

fast uses the smallest working image and coarsest position steps. accurate retains more input detail, scans more densely, and by default runs two secondary deskew passes at a bounded 640-pixel resolution. A face cannot be evaluated below the cascade's 45-pixel working window, so use accurate for small, tilted, or group portraits. Disable enableRotationRecovery when all photographs are upright and lower latency matters.

The heuristic fallback is intentionally opt-in. It helps with synthetic portraits and illustrations but can mistake symmetric objects for faces.

Detection limits #

The bundled cascade is trained for mostly frontal human faces. Accurate mode recovers modest in-plane tilt, but strong profile views, occlusion, very small faces, blur, extreme rotation, and extreme illumination can still be missed. Dartface is designed for one-shot analysis of still photographs, not real-time video.

Do not use Dartface as the only control for identity, safety, access, or liveness. Evaluate it on representative data from every intended camera, demographic, pose, and lighting condition.

Design and licensing #

The package architecture and processing pipeline are documented in docs/architecture.md.

The compact cascade parameters are generated from OpenCV's improved frontal-face LBP model. Their separate copyright, license, source revision, and research citation are recorded in THIRD_PARTY_NOTICES.md.

Built for Focale, an advanced local image editor. Discover what these packages make possible in a real creative workflow.