Siamese Instance Search for Tracking - Supplementary Material

To learn the matching function that operates on pairs of data, we use a Siamese architecture with two branches [1, 2]. The Siamese network processes the two inputs separately through individual networks that take the form of a convolutional neural network. For individual branches, we investigate two different network architectures, a small one adapted from AlexNet [5] (Figure 1a) and a very deep one inspired by VGGNet [7] (Figure 1b).