|
In this paper, we present an advanced approach to improving the performance of face detection system. While the traditional voila-Jones detector can achieve relative high accuracy, it also generates lots of false negative results in some difficult datasets especially in wild-field. To boost the recall of face detection system, we used a two-step approach. First, we make use of classical ad-boost based system to acquire coarse results; then, filter out some false negatives by the DPM-liked system which consists of some part models for face landmarks and shape model to capture the property of deformable. To train the DPM-liked models, we used a technology named data-mining to increase the efficient of training. To speed up the detection, we adopted a dynamic programming method. We show excellent performance on a dataset we gathered by ourselves and also achieve state-of-the-art performance on the less challenging BioID dataset. |
|
Keywords:face detection; deformable part model; data-mining; dynamic programming????? |
|