Generalizable Metric Network for Cross-domain Person Re-identification

TL;DR

This work tackles domain generalization in person re-identification by exploiting a key observation: domain gaps are smaller in the sample-pair feature space than in the sample-instance space. It introduces the Generalizable Metric Network (GMN), which places a lightweight Metric Network (M-Net) after the backbone and trains it on sample-pair features $f^{(sp)} = (f^{(si)}_p - f^{(si)}_q) \odot (f^{(si)}_p - f^{(si)}_q)$, enabling test-time pairwise similarity estimation. To boost generalization and discrimination, GMN adds a Dropout-based Perturbation (DP) module and a Pair-Identity Center (PIC) loss, with training driven by L_overall = L_cls + L_tri + L_gmn + lambda L_pic and DP activation after a warm-up period. Extensive experiments on Market1501, MSMT17, CUHK03-NP, and CUHK-SYSU show state-of-the-art or strongly competitive results, with ablations confirming the contribution of M-Net, DP, and PIC, and with reasonable computation using a single GPU. Overall, GMN provides an effective, scalable approach to cross-domain Re-ID by leveraging sample-pair space for robust metric learning in DG scenarios.

Abstract

Person Re-identification (Re-ID) is a crucial technique for public security and has made significant progress in supervised settings. However, the cross-domain (i.e., domain generalization) scene presents a challenge in Re-ID tasks due to unseen test domains and domain-shift between the training and test sets. To tackle this challenge, most existing methods aim to learn domain-invariant or robust features for all domains. In this paper, we observe that the data-distribution gap between the training and test sets is smaller in the sample-pair space than in the sample-instance space. Based on this observation, we propose a Generalizable Metric Network (GMN) to further explore sample similarity in the sample-pair space. Specifically, we add a Metric Network (M-Net) after the main network and train it on positive and negative sample-pair features, which is then employed during the test stage. Additionally, we introduce the Dropout-based Perturbation (DP) module to enhance the generalization capability of the metric network by enriching the sample-pair diversity. Moreover, we develop a Pair-Identity Center (PIC) loss to enhance the model's discrimination by ensuring that sample-pair features with the same pair-identity are consistent. We validate the effectiveness of our proposed method through a lot of experiments on multiple benchmark datasets and confirm the value of each module in our GMN.

Figures (4)

• Figure 1: Visualization of features in sample-instance and sample-pair spaces by t-SNE van2008visualizing. In Figure a), each dot represents a sample-instance feature, while in Figure b), each dot represents a sample-pair feature. For example, if there are two sample-instance features with $d$-dimension, i.e., the features of two images are extracted by ResNet, say $x=[x_1, \cdots, x_d] \in \mathbb{R}^{d}$ and $y=[y_1, \cdots, y_d] \in \mathbb{R}^{d}$, then the corresponding sample-pair feature can be represented as $xy=[(x_1-y_1)^2, \cdots, (x_d-y_d)^2]\in \mathbb{R}^{d}$. Both sample-instance and sample-pair features are normalized using $L_2$ normalization. Different colors denote different domains or datasets.
• Figure 2: An illustration of the proposed Generalizable Metric Network (GMN). Here we take three source domains as an example. As seen in this figure, our method yields sample-pair features based the main (backbone) network, which are leveraged to train the metric network (M-Net). Besides, the dropout-based perturbation (DP) and the pair-identity center loss ($\mathrm{L}_{pic}$) are introduced to improve the model's generalization and discrimination. It is worth noting that both backbone and M-Net are used during the test process.
• Figure 3: Comparison of different schemes on generating sample-pair features in three tasks. In this figure, "Squ" is the operation in Eq. \ref{['eq01']}. "Abs" is the absolute-value operation, i.e., $|a-b|$.
• Figure 4: Visualization of activation regions in the unseen domain. The first column is the original image, the second column is the activation map from the baseline, and the last column is the activation map from our method.