How Large-scale Training Samples Effect The Performance Of Face Detector? An Empirical Analysis

The goal of face detection is to determine whether there are any faces within a given arbitraryimage, and return the location and extent of each face in the image if one or more faces are present.Face detection gains wide practical applications, such as face recognition, video surveillance, humancomputer interface (HCI), content-based image retrieve, teleconference, digital camera and so on. Inaddition, this research topic provides useful insight for both pattern recognition and other similar targetdetection problem. Not surprisingly, face detection has been a hot issue in computer vision research inrecent years.After several decades of research, face detection has had a rapid development. Many different facedetection algorithms have been proposed and achieved very good detection results. However,the com-mon feature of most detection algorithms is that they have massive dependence of the large-scale facetraining samples which are usually constructed in a rather arbitrary manner. The internet's developmentbrings that we can easily collect large scale training samples. Large-scale face training samples can pro-vide a more diverse pattern of human faces which will help to train a face detector with a high detectionrate. At the same time, using large-scale face training samples brings some disadvantages, such as toomuch time will be cost on detector's training period, corresponding requirements on the hardware willbe increased. It has been an important issue that how the size and nature of training data affects theperformance of face detector . Study of the issue will not only help to optimize large-scale face sampleset, but also give an important revelation to sample optimization of other fields.In this paper, we empirically investigate the fundamental question of how the training set effectsthe performance of a given state of the art face detector. In particular, we construct a very large trainingset containing over 340K face images and study the effect of five common factors of variations (i.e.,lighting, expression, blurring, contrast change and noise) which may change face appearance largely.Our results show that noise factor has the most significant in?uence on the performance of the detectorwhile others (e.g., lighting, expression) are of much less importance. Based on these, we propose a newmethod which gives a certain optimization to large-scale sample set and construct an effective trainingset with much small size for face detection, without significantly reducing the performance.