face_detection_tflite #

A Dart/Flutter package that runs an on-device face analysis pipeline with TensorFlow Lite:

• Face detection (multiple SSD variants)
• 468-point face mesh
• Iris landmarks (both eyes)
• Convenience end-to-end pipeline or step-by-step access
• CPU-only, pure Dart API designed for Flutter apps on mobile and desktop

This package is a Flutter/Dart port inspired by and adapted from the original Python project patlevin/face-detection-tflite. Many thanks to the original author.

Table of Contents #

• Features
• Installation
• Quick Start
• Step-by-step Usage
• Painting Notes
• Desktop Support
• Performance Tips
• Troubleshooting
• Example
• Credits & Attribution
• License

Features #

• Face, landmark and iris detection using TensorFlow Lite models
• 468-point face mesh with normalized→image-space mapping handled for you
• Iris landmark detection for both eyes via mesh-derived ROIs
• Works on Android, iOS, macOS, Windows, and Linux
• Single-file API centered on FaceDetector
• Example app included for testing

Installation #

pub.flutter-io.cn #

dependencies:
  face_detection_tflite: ^0.1.0

Run:

flutter pub get

Quick Start #

import 'dart:typed_data';
import 'package:face_detection_tflite/face_detection_tflite.dart';

final detector = FaceDetector();

Future<void> init() async {
  // Initialize once; choose the model variant you prefer
  await detector.initialize(model: FaceDetectionModel.backCamera);
}

Future<void> analyze(Uint8List imageBytes) async {
  // All-in-one pipeline
  final result = await detector.runAll(imageBytes);

  final detections = result.detections;   // List<Detection>
  final mesh       = result.mesh;         // List<Offset> in image coordinates
  final irises     = result.irises;       // List<Offset> in image coordinates
  final size       = result.originalSize; // Size(width, height)
}

Step-by-step Usage #

If you prefer fine-grained control:

// 1) Face detections (normalized bbox + keypoints)
final detections = await detector.getDetections(imageBytes);
if (detections.isEmpty) return;

// 2) Face mesh (mapped to image coordinates)
final mesh = await detector.getFaceMesh(imageBytes);

// 3) Iris landmarks (mapped to image coordinates)
final irises = await detector.getIris(imageBytes);

// Optionally reuse intermediate outputs:
final meshFromDets = await detector.getFaceMeshFromDetections(imageBytes, detections);
final irisFromMesh = await detector.getIrisFromMesh(imageBytes, meshFromDets);

Types you will encounter:

• Detection has bbox (normalized RectF), score, and keypointsXY (normalized).
• Mesh and iris return List<Offset> in your image coordinate system.
• PipelineResult packages detections, mesh, irises, and originalSize.

Example #

A minimal example app can be placed in example/:

cd exam