Deep Learning for Person Reidentification Using Support Vector Machines

Figure 3

The framework of our proposed model. Both of positive and negative pairs are randomly selected as input images. The first to fifth layers are convolution layers and subsampling layers with Relu activation. Sixth and seventh layers are fully connected layers with 4096 neural units. The top layer is linear L2-SVM layer instead of traditional softmax layer to measure the similarity of input images.