Just as the authors state that “virtual sample generation technology belongs to the category of machine learning,” nobody ever has enough real data to train a face recognition model, and they need synthesized data. This problem is intrinsic nowadays as artificial intelligence is still weak, and major algorithms need many more samples compared to humans to learn a “simple” concept. Most algorithms fail into overfitting if not enough training data to represent the population are present. Despite numerous efforts in the past decades, facial image recognition is still challenging to use under rigorous requirements.

According to the survey, existing virtual facial image generation methods fall into three major categories: “construction of virtual face images based on the face structure, construction of virtual face images based on the idea of perturbation and distribution function of samples[,] and construction of virtual face images based on the sample viewpoint.” The first category simply utilizes the symmetry of human faces to generate virtual images similar to the source. Although simple, this type of algorithm can be very effective and easy to combine with other algorithms. If we treat the first category as biologically inspired, the second and third categories rely more on mathematical and physics models, respectively. It is notable that all the methods in the survey were handcrafted based on prior knowledge or assumptions of the human face or the imaging process; inspired by this observation, I think a fourth category is missing from the survey, which is to use machine learning techniques to generate virtual samples. Examples are generative adversarial networks (GANs) and variational autoencoders.

Finally, I do recommend section 5. While table 1 has a nice summary of 15 different algorithms surveyed, tables 3 and 4 are helpful if you need to benchmark your own algorithm.