On Low-Resolution Face Re-identification with High-Resolution-Mapping

Abstract

Low-resolution face re-identification refers to the problem of identifying if the same person’s face appears in two images: one image is low resolution (LR), e.g., from a surveillance camera, and the another one is high resolution (HR), e.g., from a government-issued identification. Research in low-resolution face re-identification has been increasing in the past few years. It can be divided into three categories: i) methods upscaling the LR image to the HR space (HR-mapping), ii) methods employing LR and HR robust features, and iii) methods learning a unified space representation. In this work, we focus on face re-identification using HR-mapping because it yields better results. Our main contribution is an experimental protocol that can be used as guideline in this task. In research, protocols are often neglected and researchers that utilize previous work on their projects have to allocate a significant amount of time to replicating inadequately described methods. We use our experimental protocol as a guideline to create a set of training and testing pairs for face re-identification using dataset VGG-Face-2. In addition, we conducted 18 experiments to validate the experimental protocol. In our experiments, we measured “d-prime” ($d'$) and the area under the ROC curve ($A_z$). We obtained: $d'=1.236$ and $A_z=0.81$ for a 14 $\times$ 14 LR pixel size set, $d'=1.900$ and $A_z=0.91$ for a 28 $\times$ 28 LR pixel size set and $d'=2.787$ and $A_z=0.97$ for a 56 $\times$ 56 LR pixel size set. We believe that our protocol can be very helpful for researchers in the field because it can be used as a set of guidelines for building comparable and replicable work of their own.