A fast and accurate cascade subspace face/eye detector on mobile devices

Mobile vision has attracted increasing research attention recently. A fast and accurate face/eye detector can enable manymobile vision applications. Most ofmobile face detections are close-range face detection. In such a scenario only a limited number of scanning windows are required. Thus the speed is less demanding but the detection accuracy is of higher importance. We further discover that for close-range face detection the covariance matrices of the face/eye class are less reliable compared to the covariance matrices of the non-face/non-eye class. Therefore, a larger weight should be assigned to the face/eye class when building the covariance mixture matrix to remove the unreliable dimensions. The proposed cascade subspace face/eye detector utilizes the focus-attention strategy in all aspects and detects eyes at precise locations at an acceptable speed. The average distance from the detected eyes to the marked eyes is 1.03 pixels only. The detected eyes are almost as accurate as the marked eyes. It takes 0.26 seconds to accurately detect both eyes on O2 XDA Flame.