Face Detection Using a 3D Model on Face Keypoints

The Support Vector Machine is a powerful learning technique that is currently lacking an efficient feature selection method that scales well to the size of the computer vision data. In this paper we bring two contributions. First, we apply a recent feature selection algorithm to optimize a differentiable version of the SVM loss with sparsity constraints. The iterative algorithm alternates parameter updates with tightening the sparsity constraints by gradually removing variables based on the coefficient magnitudes and a schedule. We use nonlinear univariate response functions to obtain a nonlinear decision boundary with feature selection and show how to mine hard negatives with feature selection. Second, we propose an approach to face detection using a 3D model on a number of detected face keypoints. The 3D model can be viewed as a simplex that fully connects the keypoints, making optimization difficult. We also propose an optimization method by that generates a set of 3D pose candidates directly by regression and verifies them with the model’s energy. Experiments on detecting the face keypoints and on face detection using the proposed 3D model show that the feature selection and nonlinear response functions dramatically improve performance and obtain state of the art face detection results on three stan-