Supplementary Material for “RGB-Infrared Cross-Modality Person Re-Identification”

This supplementary material accompanies the paper “RGB-Infrared Cross-Modality Person Re-Identification”. It includes more details of Section 4, as well as extra evaluations of our proposed deep zero-padding method. 1. Details of Counting Domain-Specific Nodes In the third paragraph of Section 4.2 in the main manuscript, we quantify the number of domain-specific nodes in the trained network in our experiments. As defined in Equation (3) in Section 3 in the main manuscript, the categorization of node types is rather strict. In the l-th layer, let η i denote the i-th node and fout(x , i, l) denote the output of η i given the network input x. Let x d1 and x (0) d2 be inputs of the whole network of domain1 and domain2, respectively. The type of node η i is defined by type(η (l) i ) =  domain1− specific, fout(x d2 , i, l) ≡ 0 domain2− specific, fout(x d1 , i, l) ≡ 0 shared, otherwise. (1) Since the identity sign is used here, the categorization condition is too strict in applications. So we relax the categorization condition for counting towards domain-specific nodes in application by setting a threshold T . The relaxed definition of node type is formulated as follows: for all x d1 and x d2 in our experiments, type(η (l) i ) =  domain1− specific, fout(x d2 , i, l) < T and fout(x (0) d1 , i, l) > T domain2− specific, fout(x d1 , i, l) < T and fout(x (0) d2 , i, l) > T shared, otherwise. (2) Because the scales of responses on feature maps differ from layer to layer, we set T = α std(x i ), where α is a proportion coefficient, x i is the output value of the i-th node in the l-th layer and std(·) is the standard deviation function. For an image channel in our experiments, we compute the average of all values in the feature map as the output of the node. We set α = 0.01 and α = 0.05 for strict and loose categorizations, respectively. The relation between the proportion of domain-specific nodes and layer depth is shown in Figure S1. Both total proportions and respective proportions of two domains are shown. With strict threshold, domain-specific nodes mainly exist in the first three layers. With loose threshold, domain-specific nodes mainly exist in the first five layers. In both cases, the network can learn more domain-specific nodes using deep zero-padding. When the threshold is loosened, the proportion of domain-specific nodes increases when using deep zero-padding, but keeps nearly unchanged when using the inputs without zero-padding. 2. Evaluation on Using Different Networks Our deep model is based on ResNet [1] as illustrated in Section 5 in the main manuscript. Deep zero-padding has shown effectiveness on ResNet-6 in our experiments. To verify whether deep zero-padding can also work with other one-stream networks, we also evaluated our method on popular architectures AlexNet [2] and VGG-16 [3]. The results are reported in Table S1. Generally, using deep zero-padding can improve the performance in most cases for all evaluated network architectures. The improvement is especially evident for ResNet-6.