Face Identification Prototypes (DL4J, OpenIMEJ, JavaCV and more)

FaceID

Kindly execute FaceID for realtime inferencing.

Face Detection

• FaceDetector.OPENCV_DL_FACEDETECTOR (default): OpenCV with prebuilt caffe face detection model
  • Configuration:
    • imageWidth:300
    • imageHeight:300
    • detectionThreshold:0.8
  • Resources:

    \OpenCVDeepLearningFaceDetector
      \res10_300x300_ssd_iter_140000.caffemodel
      \deploy.prototxt
    

• FaceDetector.OPENIMAJ_FKE_FACEDETECTOR: OpenIMAJ's FKEFaceDetector
  • Configuration
    • detectionThreshold:1.0

Face Identification

• FaceIdentifier.FEATURE_DISTANCE_VGG16_PREBUILT: Identification by highest cosine similarity between webcam image and target embeddings (prebuilt VGG16 model with VGGFACE dataset, layer fc8 as features).
  • Configuration
    • FaceFeatureProvider:VGG16FeatureProvider
    • dictDir: resources \vgg16_faces_224 detection target faces.
    • numPredicts:1 number of predictions
    • detectionThreshold:0.78
    • numSamples:3 average of top 3 per class
  • Resources:

    \vgg16_faces_224
      \person A
        \face1.jpg
        \face2.jpg
      \person B
        \face1.jpg
        \face2.jpg
    

• FaceIdentifier.CUSTOM_VGG16: Identification by inference self trained model (may refer training steps)
  • Configuration
    • numPrediction:3
• FaceIdentifier.FEATURE_DISTANCE_FACENET_PREBUILT (not stable): Identification by highest cosine similarity between webcam image and target embeddings (prebuilt InceptionResNetv1 model deployed on SKIL).
• FaceIdentifier.ZHZD: Identification by inference model trained by [email protected], mainly for testing purposes.

Dataset Preparation

• Using FaceIdentifier.FEATURE_DISTANCE: kindly invoke VGG16FaceDatasetPreperation.java to extract detection targets face images and load into \vgg16_faces_224 resource directory.