Quo Vadis RankList-based System in Face Recognition?

TL;DR

This paper revisits these RankList methods and extends them to use the logits of the state-of-the-art DaliFace network, instead of an external cohort, and shows that through a reasonable Logit-Cohort Selection (LoCoS) the performance of RankList-based functions can be improved drastically.

Abstract

Face recognition in the wild has gained a lot of focus in the last few years, and many face recognition models are designed to verify faces in medium-quality images. Especially due to the availability of large training datasets with similar conditions, deep face recognition models perform exceptionally well in such tasks. However, in other tasks where substantially less training data is available, such methods struggle, especially when required to compare high-quality enrollment images with low-quality probes. On the other hand, traditional RankList-based methods have been developed that compare faces indirectly by comparing to cohort faces with similar conditions. In this paper, we revisit these RankList methods and extend them to use the logits of the state-of-the-art DaliFace network, instead of an external cohort. We show that through a reasonable Logit-Cohort Selection (LoCoS) the performance of RankList-based functions can be improved drastically. Experiments on two challenging face recognition datasets not only demonstrate the enhanced performance of our proposed method but also set the stage for future advancements in handling diverse image qualities.

Figures (7)

• Figure 1: Tradition RL process flow. The figure illustrates the workflow of a face recognition system incorporating traditional ranklists. The dataset provides a gallery image $x_g$, a probe image $x_p$, and a cohort set, which is further divided into several images in the cohort gallery $X_{cg}$ and cohort probe $X_{cp}$. Features $\phi$ are extracted for these images firstly, where the features $\Phi$ of the cohort can be pre-computed. Then these features are used to compute similarity score lists $L_g$ and $L_p$, which are converted into two ranklists ($\gamma_g$ and $\gamma_p$). The final similarity score $s$ is computed by a dedicated rank list similarity function.
• Figure 2: LoCoS Cohort Selection Process Flow. This figure displays the flow of the proposed Logit-Cohort Selection (LoCoS) method. Logits $Z_g$ and $Z_p$ are extracted from gallery $x_g$ and probe $x_p$. Indexes are selected from the gallery logits, and applied to the probe logits, via our $S_{\mathrm{LoCoS}}$ similarity function.
• Figure 3: Traditional Approaches. This figure highlights the way of defining a cohort for traditional ranklist-based methods.
• Figure 4: ROC of Dataset SCface. This figures shows the ROC of each protocol. At FMR=$10^{-3}$ (black dot line), higher TMR indicates better performance.
• Figure 5: LosCoS-T vs. LosCoS-TB. This figure shows ROC curves on the SCface dataset, protocol FAR, when using LosCoS-T or LosCoS-TB with traditional techniques.