Mom: Mean Of Moments Feature For Person Re-Identification

Person re-identification (re-id) has drawn significant attention in the recent decade. The design of view-invariant feature descriptors is one of the most crucial problems for this task. Covariance descriptors have often been used in person re-id because of their invariance properties. More recently, a new state-of-the-art performance was achieved by also including first-order moment and two-level Gaussian descriptors. However, using second-order or lower moments information might not be enough when the feature distribution is not Gaussian. In this paper, we address this limitation, by using the empirical (symmetric positive definite) moment matrix to incorporate higher order moments and by applying the on-manifold mean to pool the features along horizontal strips. The new descriptor, based on the on-manifold mean of a moment matrix (moM), can be used to approximate more complex, non-Gaussian, distributions of the pixel features within a mid-sized local patch. We have evaluated the proposed feature on five widely used re-id datasets. The experiments show that the moM and hierarchical Gaussian descriptor (GOG) [30] features complement each other and that using a combination of both features achieves a comparable performance with the state-of-the-art methods.