How do you develop a face detector for the unconstrained environment?

With the growing of available large datasets for evaluation, face detection in recent literature has progressed rapidly. However, little research has been dedicated to develop a face detector robust to all possible variations. To address this problem, novel unconstrained datasets containing faces with more challenging variations are proposed. We notice that some recent face detectors have not been evaluated against recent unconstrained datasets. In this work, we analyse recent state-of-the-art face detectors, Headhunter and Normalized Pixel Difference (NPD), under the unconstrained conditions. We perform a detailed evaluation by proposing our evaluation protocol derived from two unconstrained datasets: (1) IARPA Janus Benchmark A (IJB-A) and (2) WIDER FACE. From our study, we find that the multiple model based detector achieves superior detection rates across most of the domains studied and requires more computational power. However, the unique combination of normalised pixel difference, soft cascade and regression trees within a face detector can achieve similar performance with much less computation.