Video Based Face Clustering

This is a framework to detect unique faces in videos.

Built with Dlib's libraries. It is possible to use different face detection and recognition algorithms.

Architecture

Input --> Preprocessing --> FaceDetection --> FaceLandmarks --> Face Alignment (Affine) --> Feature Encoder --> Clustering --> UniqueIDs

Pre-Trained Models

• OpenCV HaarCascadeClassifier: [pre-trained/haarcascade_frontalface_default.xml]
• Dlib CNN Face Detector Modelv1: pre-trained/mmod_human_face_detector.dat
• Dlib 68 Points Face Landmarks: pre-trained/shape_predictor_68_face_landmarks.dat

Installation

• git clone https://github.com/tekinengin/face-clustering-video.git
• curl http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2 > pre-trained/shape_predictor_68_face_landmarks.dat.bz2 && bzip2 -d pre-trained/shape_predictor_68_face_landmarks.dat.bz2

Parameters

• --video : Source Path
• --ctype : Face Detector(FD) Type 1: HaarCasCade, 2: HoG, 3: CNN
• --cpath : Pre-Trained FD Model Weights if any
• --ppath : Pre-Trained Feature Landmark Detector (Default Dlib-68-Points)
• --ncpu : Number of CPUs for multi-threading
• --cthreshold : Threshold for face confidence
• --resizeratio : Resize the Input with 1/resizeratio
• --pfps : Processing Fps, Example: skip videoFps / pfps frames (Default: Video Fps)
• -d : Display Option only for --ncpus 1
• -r : Saving Detected Faces and Clusters
• -e : Eye Detection Option
• -align : Face Alignment

Examples

• Clustering with Alignment (Affine) python main.py --ctype 2 --ncpu 4 --video="src/sampleVideo.mp4" --pfps 0.33 -r -align

• Clustering without Alignment :python main.py --ctype 2 --ncpu 4 --video="src/sampleVideo.mp4" --pfps 0.33 -r