Supplementary Material for Hierarchical Gaussian Descriptor for Person Re-Identification

In section 4.2 of the paper, we compared the distribution modeling of GOG to other distributions. Below, we describe the details of the compared methods. The Mean, Cov and Gauss are global distribution descriptors of pixel features within each region. The Cov-ofCov, Cov-of-Gauss and GOG are hierarchical distribution descriptors. The Cov-of-Cov uses covariance matrix in both patch and region modeling. The Cov-of-Gauss uses Gaussian for patch modeling and covariance matrix for region modeling. For a fair comparison to GOGwhich is incorporated with patch weights, we adopted the weighted pooling for all descriptors. Formally, Mean: μ′ = 1 ∑ i∈G wi ∑ i∈G wif i, Cov: Σ′ = 1 ∑ i∈G wi ∑ i∈G wi(f i − μ′)(f i − μ′)T , Gauss: P ′ = |Σ′|− 1 d+1 [ Σ′ + μ′μ′T μ′ μ′T 1 ]