Choosing a FaceNet model

Select between 128-dimensional and 512-dimensional FaceNet models for face embedding

FaceNet Android supports two different FaceNet models that produce embeddings of different dimensions. The choice between them impacts accuracy, performance, and storage requirements.

Available models

The app includes two FaceNet models in the assets folder:

• facenet.tflite - Outputs 128-dimensional embeddings
• facenet_512.tflite - Outputs 512-dimensional embeddings

Both models accept 160×160 pixel face images as input and produce embeddings that capture unique facial features.

The 512-dimensional model generally provides better accuracy for face recognition, especially with larger datasets, but requires more storage and slightly longer processing time.

Switching between models

To change the FaceNet model, you need to update two files: FaceNet.kt and DataModels.kt.

Step 1: Update the model path

In FaceNet.kt:62, modify the model file path:

FaceNet.kt

// For 128-dimensional embeddings
interpreter = Interpreter(
    FileUtil.loadMappedFile(context, "facenet.tflite"), 
    interpreterOptions
)

// For 512-dimensional embeddings
interpreter = Interpreter(
    FileUtil.loadMappedFile(context, "facenet_512.tflite"), 
    interpreterOptions
)

Step 2: Update the embedding dimension

In FaceNet.kt:34, change the embeddingDim value:

FaceNet.kt

// For facenet.tflite
private val embeddingDim = 128

// For facenet_512.tflite
private val embeddingDim = 512

Step 3: Update the database schema

In DataModels.kt:18-21, update the @HnswIndex dimensions:

DataModels.kt

@Entity
data class FaceImageRecord(
    @Id var recordID: Long = 0,
    @Index var personID: Long = 0,
    var personName: String = "",
    // Update dimensions to match your chosen model
    @HnswIndex(
        dimensions = 512,  // Change to 128 for facenet.tflite
        distanceType = VectorDistanceType.COSINE,
    ) var faceEmbedding: FloatArray = floatArrayOf(),
)

Changing the embedding dimension requires rebuilding the app and clearing existing data. The ObjectBox database schema will change, and existing face records will be incompatible.

Performance comparison

Model	Embedding Size	Accuracy	Inference Time	Storage per Face
facenet.tflite	128-dim	Good	~40-60ms	512 bytes
facenet_512.tflite	512-dim	Better	~50-70ms	2048 bytes

For most use cases with moderate dataset sizes (under 1000 faces), the 512-dimensional model provides the best balance of accuracy and performance.

Model source

Both models are sourced from the deepface library and converted to TensorFlow Lite format with FP16 optimization. See the source README for conversion scripts.