Abstract: A face detection method includes scaling an input image to images of various sizes according to certain proportions by means of an image pyramid; passing the resultant images through a first-level network in a sliding window manner to predict face coordinates, face confidences, and face orientations; filtering out the most negative samples by confidence rankings and sending the remaining image patches to a second-level network. Through a second-level network, filtering out non-face samples; applying a regression to obtain more precise position coordinates and providing prediction results of the face orientations. Through an angle arbitration mechanism, combining the prediction results of the preceding two networks to make a final arbitration for a rotation angle of each sample, rotating each of the image patches upright according to the arbitration result made by the angle arbitration mechanism and sending to a third-level network for fine-tuning to predict positions of keypoints.

Abstract: A face detection method includes scaling an input image to images of various sizes according to certain proportions by means of an image pyramid; passing the resultant images through a first-level network in a sliding window manner to predict face coordinates, face confidences, and face orientations; filtering out the most negative samples by confidence rankings and sending the remaining image patches to a second-level network. Through a second-level network, filtering out non-face samples; applying a regression to obtain more precise position coordinates and providing prediction results of the face orientations. Through an angle arbitration mechanism, combining the prediction results of the preceding two networks to make a final arbitration for a rotation angle of each sample, rotating each of the image patches upright according to the arbitration result made by the angle arbitration mechanism and sending to a third-level network for fine-tuning to predict positions of keypoints.