Face detection models

​

Available face detectors

​

Mediapipe BlazeFace (short range)

• Optimized for detecting faces within 2 meters of the camera
• Fast inference speed suitable for real-time detection
• Provides face bounding boxes and key facial landmarks
• Part of Google’s Mediapipe framework

​

Google MLKit Face Detection

• High-level API with automatic model management
• Detects faces in various orientations
• Provides facial landmarks (eyes, nose, mouth, etc.)
• Returns face contours and bounding boxes
• No manual model loading required

• Easier integration with automatic updates
• More robust detection across different angles
• Better support for edge cases

​

Choosing between detectors

@Module
@ComponentScan("com.ml.shubham0204.facenet_android")
class AppModule {

    private var isMLKit = true  // Set to false for Mediapipe

    @Single
    fun provideFaceDetector(context: Context): BaseFaceDetector = if (isMLKit) {
        MLKitFaceDetector(context)
    } else {
        MediapipeFaceDetector(context)
    }
}

​

When to use Mediapipe BlazeFace

• You want more control over the face detection model
• You need the smallest possible APK size
• You’re targeting close-range face detection scenarios
• You want to customize the TFLite model or delegate options

​

When to use MLKit

• You prefer a higher-level API with less manual configuration
• You want automatic model updates from Google
• You need robust detection across various face angles
• You want additional features like smile detection and eye-open detection

​

Mediapipe BlazeFace implementation

• Face bounding box coordinates
• Detection confidence score
• Facial landmark positions (if available)

​

MLKit implementation

• Face bounding box coordinates
• Detection confidence score
• Facial landmarks (eyes, nose, mouth, ears)
• Face contours
• Head rotation angles (Euler X, Y, Z)

​

Face detection workflow

1. Receive input: Accept a bitmap image or camera frame
2. Preprocessing: Convert to the required format (if needed)
3. Detection: Run face detection inference
4. Post-processing: Convert detector-specific results to FaceDetectionResult
5. Cropping: Extract face regions based on bounding boxes
6. Return results: Provide cropped face images for FaceNet embedding

​

Performance comparison

​

External resources

​

Mediapipe BlazeFace

• Mediapipe Face Detector documentation
• BlazeFace short-range model details
• Mediapipe Face Detection guide

​

Google MLKit

• MLKit Face Detection documentation
• MLKit Android guide
• Face Detection API reference

​

Integration with FaceNet

1. Face detection extracts bounding boxes
2. Faces are cropped to the detected regions
3. Cropped faces are resized to 160x160 pixels
4. Resized faces are fed to FaceNet for embedding generation
5. Embeddings are stored in the ObjectBox vector database
6. Cosine similarity is used to match faces during recognition