ZhengRui / cardea

context aware visual privacy control system

Cardea is a system in which user can set his/her privacy preference based on context, which includes location, surrounding scene and persons. When captured by an android camera, user's privacy preference will be enforced if context is matched. Interactive hand gestures can be used by user to overwrite his/her privacy setting temporarily.

Check demo video for the usage of the system.

Dataflow

Setup

• server:

  • modify path in server/_init_paths.py to be the correct path of py-faster-rcnn
  • modify path in server/serv.py in each worker to be the correct path of model used by this worker

• jni libraries:

  • modify paths in androidapp/jnisrc/fdlib/build.sh and build the face detection and alignment library by ./build.sh
  • build androidapp/jnisrc/caffe_android_lib

• after all model paths and jni libraries and android app set up correctly, cd server; python serv.py and use the android app as demo shows

Models

• scene classification

  • grouping, augmentation
  • feature extraction, finetune
  • merging feature extractor and classifier to a single model

• gesture recognition

  • data preparation, augmentation
  • train gesture recognizer end to end

• face recognition:

  • openface 0.2.0 post
  • A Lightened CNN for Deep Face Representation
  • face alignment