To face or not to face: Towards reducing false positive of face detection

We tackle the problem of reducing the false positive rate of face detectors by applying a classifier after the detection step. We first define and study this post classification problem. To this end, we first consider the multiple-stage cascade structure which is the most common face detection architecture. Here, each cascade stage aims to solve a binary classification problem, denoted the Face/non-Face (FnF) problem. In this context, the post classification problem can be considered as the most challenging FnF problem, or the Hard FnF (HFnF) problem. To study the HFnF problem, we propose HFnF datasets derived from the recent face detection datasets. A baseline method utilizing the GIST features and Support Vector Machine (SVM) classifier is also proposed. In our evaluation, we found that it is possible to further improve the face detection performance by addressing the HFnF problem.