Cluster and Aggregate: Face Recognition with Large Probe Set Supplementary Material

To train the fusion network F which is comprised of SIM, CN and AGN, we set the batch size to be 512. We take the pretrained model E, which is IResNet-101 [2], trained on WebFace4M [15] with ArcFace loss [2] and freeze it without further tuning. For training CAFace, the number of images per identity N is randomly chosen between 2 and 16 during each step of training, and we take two sets per identity. The intermediate feature for the Style Input Component (SIM) is taken from the block 3 and 4 of the IResNet-101. The number of clusters in CN is varied in the ablation studies and fixed to be 4 for subsequent experiments.