Face-Recognition

• To run the application, clone the repository and build it using CMake
• Should satisfy all dependencies
• Download the weight files here models
• Put .jpg images here faceImages
• 1 image should contain only 1 face and names of all files should be unique
• Name of the image file is assigned as the faceId to the face detected in the image
• No Need to crop the .jpg images, application will first detect faces in it and then store the face embeddings from face images
• Rename you video as "test_video.mp4" and move it here video

How it works

• Scans the faceImages for faces in images, extracts 512 feature vector for each image and stores them in memory
• Reads the specified input video stream frame-by-frame, be it a camera device or a video file
• The application deploys 2 models[face detection and face embeddings] and runs them in synchronous manner
• An input frame is processed by the face detection model to predict face bounding boxes
• Face images are created using the face bounding boxes and are sent to face embedding model to provide a 512 feaure vector of each face
• Face matching is performed using the feature vectors generated at run time and the ones generate from database
• Dot product/cosine distance is calculated to assign FaceIDs, if no match is found "unknown" ID is assigned to that face

Models Used

• face-detection-retail-0005 to detect faces in a frame
• landmarks-regression-retail-0009 to predict landmarks in a face Image
• SphereFace [converted to IR] so as to use with INFERENCE_ENGINE

Dependencies

• Opencv[4.5.1 and above] built with InferenceEngine backend
• dlib (for correlation tracker)

Test Results

Things to Do

• Face Detection
• Face Landmarks Detection [Added but not using currently]
• Face Recognition
• Face Alignment before recognition using landmarks
• Add Tracker and maintain state of each recognized face
• Perform face detection, landmarks detection and recognition in asynchronous manner
• Age Detection
• Gender Detection
• Emotion Detection