For covariance-based image descriptors, taking into account the curvature ofthe corresponding feature space has been shown to improve discriminationperformance. This is often done through representing the descriptors as pointson Riemannian manifolds, with the discrimination accomplished on a tangentspace. However, such treatment is restrictive as distances between arbitrarypoints on the tangent space do not represent true geodesic distances, and hencedo not represent the manifold structure accurately. In this paper we propose ageneral discriminative model based on the combination of several tangentspaces, in order to preserve more details of the structure. The model can beused as a weak learner in a boosting-based pedestrian detection framework.Experiments on the challenging INRIA and DaimlerChrysler datasets show that theproposed model leads to considerably higher performance than methods based onhistograms of oriented gradients as well as previous Riemannian-basedtechniques.
展开▼
