Qualitative Evaluation of Face Embeddings Extracted From well-known Face Recognition Models

This paper demonstrates a qualitative evaluation/comparison of face embeddings extracted from deep learning models, such as VGG-Face, Dlib, and OpenFace, on a face discrimination task. While conducting experiments, each of linear SVM (support vector machine) classifier, Euclidean distance, and Cosine distance algorithms was utilized to compare/analyze face vectors (embeddings) extracted from those 3 deep learning models. This resulted in 9 overall combinations of face recognition techniques to be compared. To implement a fair comparison of these 9 combinations, first, training and test datasets were gathered; these datasets were made up of complete frontal and well-cropped (NxN sized) face images of 33 persons (mostly Asian), with at least 10 different face images for each person. Then, face images in the training dataset were introduced into deep learning models to extract face vectors from them. Next, these face vectors were stored in a local directory as a reference database (to be used with Euclidean and Cosine distance methods) and were used to train SVM classifier. Subsequently, these face vectors were utilized to classify (recognize) face images (vectors) from the test dataset. As experiments proved, the best face recognition technique amongst 9 combinations was Dlib based face recognition model (with SVM classifier combined) as it showed the highest rate to distinguish people from each other.Although, this research work does not bring novelty to the domain, it took an effort to evaluate/compare well-known deep face models performances on Asian faces (mostly) and choose the best one to utilize as a basis for door access control application.