Boosting Speed and Accuracy in Deformable Part Models for Face Image in the Wild

Face detection using part based model becomes a new trend in Computer Vision. Following this trend, we propose an extension of Deformable Part Models to detect faces which increases not only precision but also speed compared with current versions of DPM. First, to reduce computation cost, we create a lookup table instead of repeatedly calculating scores in each processing step by approximating inner product between HOG features and weight vectors. Furthermore, early cascading method is also introduced to boost up speed. Second, we propose new integrated model for face representation and its score of detection. Besides, the intuitive non-maximum suppression is also proposed to get more accuracy in detecting result. We evaluate the merit of our method on the public dataset Face Detection Data Set and Benchmark (FDDB). Experimental results shows that our proposed method can significantly boost 5.5 times in speed of DPM method for face detection while achieve up to 94.64% the accuracy of the state-of-the-art technique. This leads to a promising way to combine DPM with other techniques to solve difficulties of face detection in the wild.