Computer Vision and1 Intelligence Group, IIT Madras

We implement an experimental setup with face detection and recognition. This has been used for our purposes with the following aims:

• Swapping multiple detectors and feature extractors for facenet.
• Multi GPU and distributed support
• Frozen graph support with quantisation.

Primarily, we use this in two use cases:

• High accuracy: SSD or FRCNN detectors with Inception-Resnet feature extractors.
• CPU optimised FPS: SSDlite mobilenet with mobilenet V2 extractors (this is covered in getting started).

We have experimented with multiple classifiers for error free student identification.

• Getting Started
• Model Zoo
• Installing Dependencies

• David Sandberg's Implementation
• Retraining Facenet
• Recognition with Facenet
• Facenet Wiki

• Object Detection Pipeline

• TF-Estimator based scalable train file.
• SSDLite based detector
• Mobilenet models for facenet
• Angular, Focal and triplet losses.
• Inference on Singular Videos.
• DALI, Tensor RT for faster inference.
• S3D support for detection.
• Experiments with weight tying.
• Results Section
• Take a look at https://github.com/alexattia/ExtendedTinyFaces for large scale face detection.

• Python 3.4+
• Tensorflow 1.6+
• Opencv 3.3.1+

Image -> FaceDetection -> CroppedFace -> FaceRecognition -> Descriptor(128D) -> FaceClassifier -> Name

TensorFlow implementation of the face recognizer described in the paper "FaceNet: A Unified Embedding for Face Recognition and Clustering". Ref. https://github.com/davidsandberg/facenet