Supplementary of Multi-scale Deep Learning Architectures for Person Re-identification

Multi-scale-A layer (Fig. 1), analyses the data stream with the size 1 × 1, 3 × 3 and 5 × 5 of receptive field. Furthermore, in order to increase both depth and width of this layer, we split the filter size of 5 × 5 into two 3 × 3 streams cascaded (i.e. stream-4 and stream-3 in Tab 1 and Fig. 1). The weights of each stream are also tied with the corresponding stream in another branch. Such a design art is, in general, inspired by, and yet different from the inception architectures [11, 12, 10]. The key difference lies in the weights which are not tied between any two streams from the same branch, but are tied between the two corresponding streams of different branches. Reduction layer (Fig. 2) further passes the data stream in multi-scale, and halves the width and height of feature maps, which should be, in principle, reduced from 78× 28 to 39 × 14. We thus employ Reduction layer to gradually decrease the size of feature representations as illustrated in Tab 1 and Fig. 2, in order to avoid representation bottlenecks. Here we follow the design principle of “avoid representational bottlenecks” [12]. In contrast to directly use max-pooling layer for decreasing feature map size, our ablation study shows that the Reduction layer, if replaced by max-pooling layer, will leads to more than 10% absolute points lower than the reported results of Rank-1 accuracy on CUHK01 dataset. Again, the weights of each filter here are also tied for paired streams. Multi-scale-B layer (Fig. 3) serves as the last stage of highlevel features extraction for the multiple scales of 1 × 1, 3 × 3 and 5 × 5. Besides splitting the 5 × 5 stream into two 3 × 3 streams cascaded (i.e. stream-4 and stream-3 in Tab 1 and Fig. 3). We can further decompose the 3 × 3 C-filters into one 1 × 3 C-filter followed by 3 × 1 C-filter [10]. This leads to several benefits, including reducing the computation cost on 3 × 3 C-filters, further increasing the