Large-Scale Face Detection in the Wild: A Comparison Study

Abstract

Various applications like face analysis, recognition, reidentification exist where the use of Face Detection is necessary as their preprocessing algorithm in the pipeline. There have been extensive studies done in the domain of Face Detection in the past, and various robust algorithms have been proposed and evaluated on different datasets. Such techniques are also deployed in various applications. Although it may seem that this domain is very old and much work must have been done in it, there is still room for improvement. Previous studies have targeted issues like facial poses, expressions, scales of images and occlusions, and have achieved good accuracy. In recent years, work on advanced issues like low-resolution images, usage of proposed anchors, scale-invariance of models, minimization of model size, have been explored and various solutions have been proposed. The proposed research work intends to experiment and evaluate various object detection techniques, designed specifically for frontal faces, on a dataset of images containing faces. The dataset is prepared using the images from a live feed of a news channel and contains various illumination, scale, quality variations making the dataset complex enough to test the techniques on. Another dataset of face images containing medical face masks is used, to evaluate how good well such models perform in presence of occlusions like surgical masks. These models tested on these datasets are evaluated based on the Mean Average Precision (mAP) metric which uses Intersection over Union (IoU) to calculate the True Positive, False Positive, True Negative, and False Negative for accurate precision calculation. Results are compiled and compared to finalize which type of model works best for the datasets under observation. .