Skip to main content
The FaceNet class provides face embedding generation using a TensorFlow Lite model. It converts face images into 512-dimensional embedding vectors for face recognition tasks.

​
Constructor

FaceNet(
    context: Context,
    useGpu: Boolean = true,
    useXNNPack: Boolean = true
)
​
context
Context
required
Android application context for loading the TFLite model from assets
​
useGpu
Boolean
default:"true"
Enable GPU acceleration if supported by the device. Falls back to CPU with 4 threads if GPU is not available
​
useXNNPack
Boolean
default:"true"
Enable XNNPACK delegate for optimized inference on CPU

​
Methods

​
getFaceEmbedding

Generates a 512-dimensional face embedding from a cropped face image. 
suspend fun getFaceEmbedding(image: Bitmap): FloatArray
​
image
Bitmap
required
Cropped face image to generate embedding from. Will be automatically resized to 160x160 pixels
​
return
FloatArray
A 512-dimensional embedding vector representing the face features

​
Configuration

The FaceNet model has the following fixed configurations:
​
imgSize
Int
default:"160"
Input image size (160x160 pixels). Images are automatically resized to this dimension
​
embeddingDim
Int
default:"512"
Output embedding dimension. The model generates 512-dimensional vectors
​
modelFile
String
default:"facenet_512.tflite"
TFLite model file loaded from the assets folder

​
Usage example

// Initialize FaceNet with GPU acceleration
val faceNet = FaceNet(
    context = applicationContext,
    useGpu = true,
    useXNNPack = true
)

// Generate embedding from a cropped face
val embedding = faceNet.getFaceEmbedding(croppedFaceBitmap)

// Store embedding in database
imagesVectorDB.addFaceImageRecord(
    FaceImageRecord(
        personID = personID,
        personName = personName,
        faceEmbedding = embedding
    )
)

​
Implementation details

The class automatically:
  • Resizes input images to 160x160 pixels using bilinear interpolation
  • Normalizes pixel values from [0, 255] to [0, 1]
  • Detects GPU compatibility and enables GPU delegate when available
  • Falls back to CPU with 4 threads if GPU is unavailable
  • Enables NNAPI acceleration for compatible devices
Source: domain/embeddings/FaceNet.kt:25

Build docs developers (and LLMs) love

Get started for freeTalk to us