Log In Sign Up

Feature-Distribution Perturbation and Calibration for Generalized Person ReID

by   Qilei Li, et al.

Person Re-identification (ReID) has been advanced remarkably over the last 10 years along with the rapid development of deep learning for visual recognition. However, the i.i.d. (independent and identically distributed) assumption commonly held in most deep learning models is somewhat non-applicable to ReID considering its objective to identify images of the same pedestrian across cameras at different locations often of variable and independent domain characteristics that are also subject to view-biased data distribution. In this work, we propose a Feature-Distribution Perturbation and Calibration (PECA) method to derive generic feature representations for person ReID, which is not only discriminative across cameras but also agnostic and deployable to arbitrary unseen target domains. Specifically, we perform per-domain feature-distribution perturbation to refrain the model from overfitting to the domain-biased distribution of each source (seen) domain by enforcing feature invariance to distribution shifts caused by perturbation. Furthermore, we design a global calibration mechanism to align feature distributions across all the source domains to improve the model generalization capacity by eliminating domain bias. These local perturbation and global calibration are conducted simultaneously, which share the same principle to avoid models overfitting by regularization respectively on the perturbed and the original distributions. Extensive experiments were conducted on eight person ReID datasets and the proposed PECA model outperformed the state-of-the-art competitors by significant margins.


Federated and Generalized Person Re-identification through Domain and Feature Hallucinating

In this paper, we study the problem of federated domain generalization (...

Label Distribution Learning for Generalizable Multi-source Person Re-identification

Person re-identification (Re-ID) is a critical technique in the video su...

Debiased Batch Normalization via Gaussian Process for Generalizable Person Re-Identification

Generalizable person re-identification aims to learn a model with only s...

Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification

Learning generic and robust feature representations with data from multi...

Dual Distribution Alignment Network for Generalizable Person Re-Identification

Domain generalization (DG) serves as a promising solution to handle pers...

On Calibration and Out-of-domain Generalization

Out-of-domain (OOD) generalization is a significant challenge for machin...

Resisting Out-of-Distribution Data Problem in Perturbation of XAI

With the rapid development of eXplainable Artificial Intelligence (XAI),...

1 Introduction

Person Re-identification (ReID) aims to identify the images of the same pedestrians captured by non-overlapping cameras at different times and locations. It has achieved remarkable success when both training and testing are performed in the same domains  Li et al. (2018c); Zheng et al. (2019); Zhang et al. (2020). However, the widely held i.i.d. assumption does not always hold in real-world ReID scenarios due to significantly diverse viewing conditions at different locations of biased distributions at different camera views, and more generally across different application domains. As a result, a well-trained model will degrade significantly when applied to unseen new target domains Luo et al. (2020); Choi et al. (2021); Wei et al. (2018). To that end, Domain Generalization (DG) Zhou et al. (2021, 2020); Mahajan et al. (2021), which aims at learning a domain-agnostic model, has drawn increasing attention in the ReID community. It is a more practical and challenging problem, which requires no prior knowledge about the target test domain to achieve “out-of-the-box” deployment.

Recent attempts on generalized ReID aim to prevent models from overfitting to the training data in source domains from either a local perspective by manipulating the data distribution of each domain, or in a global view to represent the samples of all domains in a common representational space. The local-based methods Dai et al. (2021); Yu et al. (2021); Jin et al. (2020a); Jia et al. (2019) are usually implemented by feature perturbation and/or normalization, as shown in Figure 1 (a). However, the perturbed distributions constructed from the original data of a single source domain is subject to subtle distribution shift and also domain biased. On the other hand, the global-based approaches  Choi et al. (2021); Ang et al. (2021); Zhou et al. (2020); Zhang et al. (2021) aim to align the feature distributions of multiple domains so that the per-domain data characteristic (i.e.

, mean and variance of the data distribution which is assumed to be a Gaussian distribution) is ignored when representing images of different domains, as illustrated in Figure 

1 (b). They often explicitly pre-define a target distribution to be aligned towards, or implicitly learn a global consensus by training a single model with data of all the source domains. However, even the domain gap is reduced by such a global regularization from restricted ‘true’ distributions, the learned representations are inherently domain-biased toward the consensus of the multiple seen training domains rather than the desired universal distribution scalable to unseen target domains given the number of domains available for training is always limited.

In this work, we present a Feature-Distribution Perturbation and Calibration (PECA) model to accomplish generalized ReID with the objective to learn more generalizable discriminative representations for model deployment to unseen target domains. This is achieved by regularizing model training simultaneously with local distribution perturbation and global distribution calibration, as depicted in Figure 1

(c). Specifically, on the one hand, as each source domain usually depicts limited numbers of pedestrians under certain scenarios, simply training from such data will lead to overfitting to the domain-specific inherently domain biased distribution, which harms the model’s generalizability. To address this issue, we introduce local perturbation module to diversify the feature distribution based on a perturbing factor estimated per domain, which enables the model to be more invariant to distribution shifts. On the other hand, despite the unpredictable distribution gaps between different ReID data domains due to the undesirable scenario-sensitive information embedded in images specific to each domain,

e.g., the background, we consider that the features derived from different independent domains should share a high proportion of information as the universally applicable explanatory factors for domain-independent identity discrimination. In this regard, we propose to simultaneously calibrate the feature distributions across all the source domains, so to eliminate the domain-specific data characteristics in feature representations that are potentially caused by the identity-irrelevant redundancy. Both the proposed local perturbation and global calibration modules reinforce the same purpose of regularizing the model training, but they are devised in different hierarchies and complementary to each other. Different from the existing methods which consider only partially from the local or global perspectives, our method handles both to promote the model in learning domain-agnostic representations.

Figure 1: Illustration of three training schemes in domain generalized ReID. The ‘Universal’ is ideally the distribution for any new target domains. The source domains are differentiated by different colors, and the perturbed distributions share the same color with the corresponding original. The number indicates the characteristics of that distribution, and similar value means a smaller domain gap, vice versa. The proposed PECA model simultaneously conducts local perturbation and global calibration to eliminate domain bias for learning a domain-agonistic representation.

Contributions of this work are three-fold: (1) To our best knowledge, we make the first attempt to exploit jointly the local feature-distribution perturbation and the global feature-distribution calibration for improving the model’s generalizability to arbitrary unseen domains while maintaining its discrimination. (2) We formulate a local perturbation module (LPM) to diversify per-domain feature distribution so to refrain the model from overfitting to each source domain, and a global calibration module (GCM) to further eliminate domain bias by aligning the distribution of multiple source domains. We simultaneously regularize both to strike the optimal balance between these two competing objectives. (3). Extensive experiments verify the superior generalizability of the proposed PECA model over the state-of-the-art DG models on a wide range of ReID datasets by a notable margin, e.g., the mAP is improved absolutely by 5.8% on Market1501 and the Rank-1 is by 6.0% on MSMT17.

2 Related Works

Generalized Person ReID. Person ReID aims to match the same identity across disjoint cameras. However, a well-trained ReID model always significantly degrades when evaluated on novel unseen domains, caused by the domain bias between training and testing data. To obtain a robust model in achieving “out-of-the-box” deployment, recent attempts on generalized ReID avoids models from overfitting to the source domains by either local domain manipulation or global cross-domain alignment to extract domain-invariant features more resistant to domain bias.

Local domain data manipulation.

It is easy to train separate local models with labeled samples respective to each source domain, with subsequently a model aggregation Dai et al. (2021); Yu et al. (2021). However, these local models would overfit to the corresponding source domain, while losing generalizability to others. A natural way to solve this problem is either to diverse the local training samples for learning a knowledgeable model, or to eliminate biased information within each source domain for learning a domain-unbiased model. Both solutions fall into the category of local domain manipulation, which alters the data distribution in a per-domain manner. For data diversification, the most intuitive approach is to perform augmentation, on either the raw image Chen et al. (2020) or feature spaces Fu et al. (2019). For eliminating local domain bias, normalization, which regulates the data distribution based on the data statistics, has been widely studied recently Jin et al. Jin et al. (2020a) introduced instance normalization (IN) for restituting the style component out of an ID representation. Jia et al. Jia et al. (2019)

combined batch normalization with instance normalization in a unified architecture to achieve content and style unification. However, these local diagrams consider only per-domain information during feature perturbation, and is still subject to subtle distribution shifts. In this work, we propose to perturb the per-domain feature distribution to empower the model to be agnostic to holistic domain shift. The complementary regularization provided by the global distribution calibration remedy helps the learned model being invariant against both perturbed distribution shift and real domain gap, so to extract generic yet discriminative representation for any unseen domain.

Global distribution calibration.

In contrast to the local approaches, methods based on global distribution calibration consider the cross-domain association by learning a shared representational space for all domains. These methods are built based on a straightforward assumption that source invariant features are also invariant to any unseen target domains Li et al. (2018b). In this spirit, DEX Ang et al. (2021)

dynamically performed e space expansion towards the direction of a zero-mean normal distribution with a covariance matrix estimated from the corresponding domain. Recent works 

Zhao et al. (2021); Choi et al. (2021) took the idea of meta-learning with the aim of “learning to generalize” by randomly splitting available source domains into meta-training and meta-testing sets, to mimic real-world deployment scenarios. Such a scheme implicitly aligns the cross-domain feature distributions to a shared space by randomly setting the alignment target, i.e., the meta-testing set. Zhang Zhang et al. (2021) proposed learning causal invariant feature by disentangling ID-specific and domain-specific factors for all the training samples from all the source domains, which enables the disentangled feature to well-preserved ID information while sharing the same feature space for all the domains. However, even aligning among multiple ‘real’ source domains can reduce the domain gap, the learned representations are still biased towards the consensus of the limited seen training domains, instead of the desired universal distribution scalable to unseen target domains. In this work, we propose to associate global alignment with local perturbation to achieve hierarchical regularization to avoid the model from overfitting to the source domains, so to learn domain-agnostic representations.

Data Augmentation. The conventional paradigm of data augmentation is to diversify data. GANs Goodfellow et al. (2014) have also been extensively explored to generate new data samples. Yang et al. Yang et al. (2021) designed an image augmentation module which helps the network to learn domain-invariant representation by distilling information learned from the augmented samples to the teacher network. More recently, feature augmentation has emerged for semantic transformations. DeepAugment Hendrycks et al. (2021) perturbed features via stochastic operations by forwarding images through a pre-trained image-to-image model, to generate semantically meaningful and diverse samples. Li et al. Li et al. (2021) discovered that embedding white Gaussian noise in high-dimensional feature space provides substantive statistics reflective of cross-domain variability. Li et al. Li et al. (2022) proposed to model the feature uncertainty with a multivariate Gaussian distribution to perturb hierarchical features to diversify the feature space. In this work, we explore feature distribution augmentation in each source domain to achieve per-domain feature distribution diversification rather than diversifying the data, with the objective of making the model invariant to per-domain holistic shift in order to avoid model overfitting in each source domain.

Distribution Alignment. The idea of distribution alignment aims to minimize the feature discrepancy between source and target domains. However, it is impossible for DG to explicitly perform such “target-oriented” alignment due to the absence of target domains during model training. With a straightforward assumption that features which are invariant to the source domain shift should also be invariant to any unseen target domain Li et al. (2018b), DG approaches share the spirit to minimize the discrepancy among source domains to achieve distribution alignment. There are a wide variety of statistical metrics available for minimizing, such as Euclidean distance and -divergences. In this regard, Li et al. Li et al. (2020)

proposed to minimize the KL divergence of source domain features with a Gaussian distribution. Several researchers achieved distribution alignment by minimizing a single moment (mean or variance) 

Muandet et al. (2013); Hu et al. (2020) or multiple moments  Erfani et al. (2016); Ghifary et al. (2016) calculated over a batch of source domain samples through either a projection matrix Ghifary et al. (2016) or a non-linear deep network Jin et al. (2020b). Li et al. Li et al. (2018b) minimized the MMD distance by aligning the source domain feature distributions with a prior distribution via adversarial training Goodfellow et al. (2014). In this paper, the proposed global distribution calibration operates on the same principle to align the source domains in learning a domain-agnostic model. Differently, we tailor the alignment objective for person ReID considering that all samples are depicting pedestrians, rather than predefine a deterministic distribution to align, e.g., Gaussian or Laplace distributions. Specifically, we constructed a common feature space upon the ID prototypical representations stored in a global memory bank, so as to eliminate domain-biased information.

3 Balancing Feature-Distribution Local Perturbation with Global Calibration

Given source domains , the objective of generalized ReID is to derive a domain-agnostic model which is capable of extracting domain-invariant representations for identity retrieval by a distance metric, e.g.

, Cosine similarity or Euclidean distance, for any

unseen target domain , This is inherently challenging due to the unpredictable domain gap between training and testing data.

Figure 2: Overview of the proposed Feature-Distribution Perturbation and Calibration (PECA) model. The overall objective is to derive generic feature representation by avoiding model overfitting to the source domains, which is achieved by Local Perturbation Module to enforce the learned feature invariant to per-domain distribution shifts caused by perturbation, and Global Calibration Module to align cross-domain distribution regardless of domain annotations.

3.1 Overview

In this work, we propose a Feature-Distribution Perturbation and Calibration (PECA) model to derive domain-agnostic yet discriminative ID representations. It regularizes the model training to satisfy simultaneously both local perturbation and global calibration. The local regularization is built to perform per-domain feature-distribution diversification, and the global calibration is designed to achieve cross-domain feature-distribution alignment, as shown in Figure 2. During training, for each source domain , a batch of samples is fed into the network backbone to extract the feature map . Then we perform per-domain diversification with Local Perturbation Module (LPM) as


where is the function of LPM to enable the local model to be invariant against per-domain shifts by training with the perturbed features .

The balancing Global Calibration Module (GCM) further regularizes the model learning by aligning the holistic representation (the input feature of the classifier) into a common feature space constructed from

regardless of domain label. To distinguish the holistic representation from the intermediate representation , we note it as and its perturbed counterpart as correspondingly, where

is a hyperparameter to the representation dimension. This global regularization is mathematically formulated as


where is the global regularization term aiming to align the distribution of holistic ID representations with the global distribution .

As complementary to the LPM, GCM focus on cross-domain regularization by pulling representations into a domain-agnostic space, thus empowering the generalizability of the ReID model for any unseen novel domain. With the collaboration of LPM and GCM, the PECA model can be trained with arbitrarily conventional ReID objectives in an end-to-end manner. When deployed to an unseen novel domain, a generic distance metric (e.g., Euclidean or Cosine distance) is used to measure the pairwise representational similarity between the query image against the galleries for identity retrieval.

3.2 Local Feature-Distribution Perturbation

Given an intermediate feature representation extracted from the source domain at -th layer, the objective of LPM is to perturb per-domain features to avoid local-domain overfitting. For notation clarity, we omit the layer index in the following formulations. Inspired by feature augmentation Li et al. (2021) and Instance Normalization (IN) Huang and Belongie (2017); Li et al. (2022), LPM performs perturbation by randomly substituting the transformation factors of IN. Specifically, we first calculate the channel-wise moments and for IN as


As suggested by Jin et al. (2020a), these statistical moments encode not only style information but also certain task-relevant information dedicated to ReID. Instead of discarding all of them for style bias reduction as adopted in Jia et al. (2019); Zhao et al. (2021), we propose to maintain the discrimination while increasing the local-domain data diversity by holistically shifting its distribution. This is achieved by perturbing the per-domain instance moments as


where calculate the perturbation factors, which are mathematically the standard derivation. They reflect the dispersed level of the local domain, and ensures the perturbation within a plausible range, so to avoid over-perturbation which causes model collapse, or under-perturbation which cannot provide any benefit in model learning. and varies the perturbation intensity to guarantee the diversity of perturbed features, and both are randomly sampled from a standard normal distribution. We subsequently perform feature transformation by substituting the local-domain moments as


By introducing the perturbed representation , the per-domain feature becomes more diverse so to improve the model’s generalizability against the per-domain shift.

3.3 Global Feature-Distribution Calibration

The global calibration module (GCM) is complementary to LPM by aligning the distribution of cross-domain features into a common feature space. GCM considers the association between the perturbed holistic representation and a global memory bank . Specifically, we calculate the global statistical moments and in each training iteration as


where is the prototypical feature of the -th identity in the -th domain. These global statistical moments depict a feature space shared by the prototypical representations on for all the identities. Subsequently, the holistic representations are calibrated into the joint feature space by


Here, and are the channel-wise mean and standard derivation of the perturbed representation . GCM enables the extracted features to fall into a domain-invariant space. The hierarchical regularization achieved by LPM and GCM makes the model generic in extracting domain-agnostic representations.

3.4 Training Pipeline

Learning objective.

Given the formulations of LPM and GCM, the proposed PECA can benefit from conventional learning supervisions. Specifically, the PECA model is jointly trained with a softmax cross-entropy loss and the global regularization item as


The notations and are the raw input images sampled from domain and its corresponding ID label, respectively, whilst is a one-hot distribution activated at . The function stands for the memory-based classifier Zhong et al. (2019); Zhao et al. (2021), and decides the importance of regarding the identity loss .

Memory bank update.

In each training iteration, once the network parameters are updated according to (Eq. (8)), the memory bank is then refreshed by Exponential Moving Average (EMA) as


in which is the EMA momentum. The prototypical features in the memory bank is iteratively updated with the latest corresponding ID representations. Consequently, a more discriminative feature space will be yielded by for global alignment.

4 Experiments

Figure 3: Example identity samples from different domains. Significant domain gaps are caused by the variation on nationality, illumination, viewpoints, resolution, scenario, etc.

Datasets and protocols.

We conducted multisource domain generalization on a wide-range of benchmarks, including Market1501 (M) Zheng et al. (2015), DukeMTMC (D) Zheng et al. (2017), MSMT17 (MT) Wei et al. (2018), CUHK02 (C2) Li and Wang (2013), CUHK03 (C3) Li et al. (2014), CUHK-SYSU (CS) Xiao et al. (2016), and four small datasets including PRID Hirzer et al. (2011), GRID Loy et al. (2010), VIPer Gray and Tao (2008), and iLIDs Zheng et al. (2009). The statistics of these datasets are shown in Table 1, and a few samples are visualized in Figure 3

, which reveals significant domain gaps caused by variations on nationality, illumination, viewpoint, resolution, scenario, etc. Mean average precision (mAP) and CMC accuracy on Rank-1 are adopted as evaluation metrics.

Probe Gallery
Datasets ID Img ID Img
PRID Hirzer et al. (2011) 100 100 649 649
GRID Loy et al. (2010) 125 125 900 900
VIPeR Gray and Tao (2008) 316 316 316 316
iLIDS Zheng et al. (2009) 60 60 60 60
Abbr. ID Img
Market1501 Zheng et al. (2015) M 1,501 29,419
DukeMTMC Zheng et al. (2017) D 1,812 36,411
MSMT17 Wei et al. (2018) MS 4,101 126,441
CUHK02 Li and Wang (2013) C2 1,816 7,264
CUHK03 Li et al. (2014) C3 1,467 14,097
CUHK-SYSU Xiao et al. (2016) CS 11,934 34,574
Table 1: Statistics of ReID datasets.

Implementation details.

We used ResNet50 He et al. (2016)

pre-trained on ImageNet to bootstrap our feature extractor. The batch size was set to 128, including 16 identities and 8 images for each. All images were resized to

. We randomly augmented the training data by cropping, flipping, and colorjitter. The proposed PECA was trained 60 epochs by Adam optimizor 

Kingma and Ba (2015), and we adopted the warm-up strategy in the first 10 epochs to stabilize model training. The learning rate was initialized as

and multiplied by 0.1 at 30th and 50th epoch. The momentum for the memory update was set to 0.8. The dimension of extracted representations was conventionally set to 2048. All the experiments were conducted on the PyTorch 

Paszke et al. (2017) framework with four A100 GPUs.

4.1 Comparisons to the State-Of-The-Art

Comparison under the traditional benchmark setting.

Under the existing benchmark setting Dai et al. (2021); Jin et al. (2020a); Song et al. (2019), five datasets (M+D+C2+C3+CS, as in Table 1) were used as source domains, and the generalizability was evaluated on four small-scale datasets of different domains not contributing to training (unseen), which are PRID, GRID, VIPeR, iLIDs. All the images in the source domains were used for training, without the original training or testing splits. Being consistent with existing performance evaluation protocols Jin et al. (2020a); Song et al. (2019), we performed 10-trail evaluations by randomly splitting query/gallery sets, and reported the averaged performance in Table 2, which shows the considerable superiority of the proposed PECA over the state-of-the-art (SOTA) competitors.

Method mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1
AggAlign Zhang et al. (2017) 25.5 17.2 24.7 15.9 52.9 42.8 74.7 63.8 44.5 34.9
Reptile Nichol et al. (2018) 26.9 17.9 23.0 16.2 31.3 22.1 67.1 56.0 37.1 28.0
CrossGrad Shankar et al. (2018) 28.2 18.8 16.0 9.0 30.4 20.9 61.3 49.7 34.0 24.6
Agg_PCB Sun et al. (2019) 32.0 21.5 44.7 36.0 45.4 38.1 73.9 66.7 49.0 40.6
MLDG Li et al. (2018a) 35.4 24.0 23.6 15.8 33.5 23.5 65.2 53.8 39.4 29.3
PPA Qiao et al. (2018) 45.3 31.9 38.0 26.9 54.5 45.1 72.7 64.5 52.6 42.1
DIMN Song et al. (2019) 52.0 39.2 41.1 29.3 60.1 51.2 78.4 70.2 57.9 47.5
SNR Jin et al. (2020a) 66.5 52.1 47.7 40.2 61.3 52.9 89.9 84.1 66.3 57.3
RaMoE Dai et al. (2021) 67.3 57.7 54.2 46.8 64.6 56.6 90.2 85.0 69.1 61.5
PECA (Ours) 72.2 62.7 59.4 48.4 70.1 61.2 85.7 79.8 71.9 63.0
Table 2: Comparisons with the SOTA methods under traditional setting. Best results are in bold.

Comparison under large-scale benchmark setting.

We further evaluated our model on four large-scale datasets (M+D+C3+MS) with the ‘leave-one-out’ strategy, namely taking three datasets used as source domains for model training, and one left out as an unseen target domain. Under this setting, The original train splits in the three source domains were used for training, while the test split on the unseen target domain was used for testing, same as in Zhao et al. (2021). The evaluation results in Table 3 show that PECA outperforms the SOTA competitors by a compelling margin, Specially, on the more challenging datasets CUHK03 and MSMT17 with larger domain gaps to the other datasets, all methods give relatively poorer generalization performances. In comparison, our PECA model gains greater advantage over the other methods especially on Rank-1 scores. This suggests PECA’s better scalability with greater potential in real-world deployment to different unseen target domains.

Market-1501 DukeMTMC CUHK03 MSMT17 Average
Method mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1
QAConv Liao and Shao (2020) 39.5 68.6 43.4 64.9 19.2 22.9 10.0 29.9 28.0 46.6
M3L Zhao et al. (2021) 51.1 76.5 48.2 67.1 30.9 31.9 13.1 32.0 35.8 51.9
M3L(IBN) Zhao et al. (2021) 52.5 78.3 48.8 67.2 31.4 31.6 15.4 37.1 37.0 53.5
PECA (Ours) 58.3 81.4 49.8 70.0 34.1 35.5 17.7 43.1 40.0 57.5
Table 3: Comparisons with the SOTA generalized person ReID models on large-scale datasets.

4.2 Ablation Study

Components analysis.

We investigated the effects of different components in PECA model design to study individual contributions. We trained a baseline model with only identity loss , and then incorporated it with either LPM or GCM as well as both (PECA). Table 4 shows that both the LPM and GCM are beneficial individually, and the benefits become clearer when they are jointly adopted as in the PECA model. From another perspective, it also verifies that solely considering the local or global regularization is biased, and it is non-trivial that the PECA explores both in a unified framework to learn a more generic representation.

Market-1501 DukeMTMC CUHK03 MSMT17 Average
Setting mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1
baseline 54.1 78.5 49.0 68.1 31.1 31.9 14.9 38.1 37.3 54.1
+LPM 57.9 80.4 49.4 69.4 32.7 33.2 17.7 42.8 39.4 56.5
+GCM 55.0 79.5 49.0 68.5 32.6 33.6 16.1 39.4 38.2 55.2
PECA 58.3 81.4 49.8 70.0 34.1 35.5 17.7 43.1 40.0 57.5
Setting mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1
baseline 69.1 59.0 59.0 48.4 68.9 60.1 82.5 74.5 69.9 60.5
+LPM 71.5 61.2 58.0 48.5 69.7 60.9 85.3 78.7 71.1 62.3
+GCM 69.7 59.5 59.1 48.5 69.7 60.7 85.3 78.7 71.0 61.8
PECA 72.2 62.7 59.4 48.4 70.1 61.2 85.7 79.8 71.9 63.0
Table 4: Components analysis of LPM and GCM. PECA incorporates both in a unified framework.

Discrimination and generalization trade-off.

There is a trade-off between being discriminative to the source domains, and being generalized to the target domains Zhang et al. (2018). We quantitatively assessed the proposed PECA model in this regard. The results in Table 5 indicate the baseline method fails to generalize well to the target domains but yielded compelling discrimination capacity in the source domains, which is likely due to overfitting. As a comparison, our PECA gains notable improvements in generalization ability with only slight performance drops in the source domains. This implies that PECA can effectively balance the generalization and discrimination of feature representations, so to be applied to any novel unseen domains.

Source Average Target: M Source Average Target: D
Setting mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1
baseline 58.6 73.8 54.1 78.5 63.0 77.4 49.0 68.1
LPGC 57.9 73.0 58.3 81.4 61.9 76.3 49.8 70.0
Source Average Target: C Source Average Target: MS
Setting mAP Rank-1 mAP Rank-1 mAP Rank-1 mAP Rank-1
baseline 64.2 82.3 31.1 31.9 69.9 79.3 14.9 38.1
LPGC 63.9 82.2 34.1 35.5 68.8 78.7 17.7 43.1
Table 5: Local discrimination and global generalization trade-off.

Effects of distribution perturbation on different layers.

(a) Averaged performance under large-scale setting.
(b) Averaged performance under traditional setting.
Figure 4: Effects of distribution perturbation on different layers.

We studied the effects of perturbing the input distributions of various layers in our backbone network, including the ‘Shallow’ layers (the first convolution layer and a following residual block), and ‘Deep’ layers (the last two residual blocks). The results are shown in Figure 4. It is not a surprise that perturbing the shallow layers consistently improves the performance under both the traditional and large-scale settings, as perturbations in earlier stages helps enhance the invariance of most layers to distribution shift. However, solely perturbing the deep layers exhibit distinct behaviors under different benchmark settings. This is because the training data under the traditional setting is relatively smaller with restricted diversity and perturbations in later stages tends to affect a limited part of the network that is insufficient to improve the model’s generalization ability. Based on these observations, we propose to perturb all the layers to improve the robustness of the PECA model regardless of the dataset scale.

Effects of the global calibration objective.

Traditional setting Large-scale setting
Setting mAP Rank-1 mAP Rank-1
PECA (default, ) 71.9 63.0 40.0 57.5
PECA w/o GPM 71.0 61.9 37.5 54.3
PECA w/ 71.2 62.2 39.4 56.5
PECA w/ 70.8 62.1 39.2 56.5
PECA w/ 33.9 23.9 38.5 55.5
Table 6: Effects of the global calibration objective whose importance is decided by the weight in Eq. 8. Averaged performances are reported.

The importance of the global calibration objective for avoiding the model from overfitting to source domains is determined by the hyperparameter in Eq. (8). By linearly varying from 0.1 to 100, we observed from Table 6 that moderately applying GCM (e.g., 0.1 or 1) is beneficial to PECA’s generalizability; further increasing to a larger value (e.g., 10 or 100) brings more harm than help. This is because the learning process is dominated by the calibration regularization and the model can barely learn from the identity loss, hence, the resulted feature is less discriminative. We also observed that the traditional setting is relatively more sensitive to , as it holds much less training data for learning a robust model, and a similar phenomenon is shown in Figure 4. Given the above observations, we set in practice for our PECA model.

5 Conclusions

In this work, we presented a novel Feature-Distribution Perturbation and Calibration (PECA) model to learn generic yet discriminative representation in multiple source domains generalizable to arbitrary unseen target domains for more accurate unseen domain person ReID. PECA simultaneously conducts model regularization on local per-domain feature-distribution and global cross-domain feature-distribution in order to learn a better domain-invariant feature space representation. Benefited from the diverse features synthesized by local perturbation, PECA expands per-domain feature distribution to enable more robust to domain shifts. From the global calibration, feature distributions of different domains are represented and holistically referenced in a shared feature space with their domain-specific data characteristics (i.e., mean and variance of feature distributions) being ignored, resulting in higher model generalizability. Experiments on extensive ReID datasets show the performance advantages of the proposed PECA model over a wide range of state-of-the-art competitors. Extensive ablation studies further provided in-depth analysis of the individual components designed in PECA model.


  • [1] E. P. Ang, L. Shan, and A. C. Kot (2021) DEX: domain embedding expansion for generalized person re-identification. In BMVC, Cited by: §1, §2.
  • [2] F. Chen, N. Wang, J. Tang, D. Liang, and H. Feng (2020) Self-supervised data augmentation for person re-identification. Neurocomputing 415, pp. 48–59. Cited by: §2.
  • [3] S. Choi, T. Kim, M. Jeong, H. Park, and C. Kim (2021) Meta batch-instance normalization for generalizable person re-identification. In CVPR, Cited by: §1, §1, §2.
  • [4] Y. Dai, X. Li, J. Liu, Z. Tong, and L. Duan (2021) Generalizable person re-identification with relevance-aware mixture of experts. In CVPR, Cited by: §1, §2, §4.1, Table 2.
  • [5] S. Erfani, M. Baktashmotlagh, M. Moshtaghi, X. Nguyen, C. Leckie, J. Bailey, and R. Kotagiri (2016) Robust domain generalisation by enforcing distribution invariance. In IJCAI, Cited by: §2.
  • [6] Y. Fu, Y. Wei, G. Wang, Y. Zhou, H. Shi, and T. S. Huang (2019) Self-similarity grouping: a simple unsupervised cross domain adaptation approach for person re-identification. In ICCV, Cited by: §2.
  • [7] M. Ghifary, D. Balduzzi, W. B. Kleijn, and M. Zhang (2016) Scatter component analysis: a unified framework for domain adaptation and domain generalization. IEEE TPAMI 39 (7). Cited by: §2.
  • [8] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio (2014) Generative adversarial nets. NeurIPS. Cited by: §2, §2.
  • [9] D. Gray and H. Tao (2008) Viewpoint invariant pedestrian recognition with an ensemble of localized features. In ECCV, Cited by: §4, Table 1.
  • [10] K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In CVPR, Cited by: §4.
  • [11] D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo, R. Desai, T. Zhu, S. Parajuli, M. Guo, et al. (2021) The many faces of robustness: a critical analysis of out-of-distribution generalization. In ICCV, Cited by: §2.
  • [12] M. Hirzer, C. Beleznai, P. M. Roth, and H. Bischof (2011) Person re-identification by descriptive and discriminative classification. In Scandinavian conference on Image analysis, Cited by: §4, Table 1.
  • [13] S. Hu, K. Zhang, Z. Chen, and L. Chan (2020) Domain generalization via multidomain discriminant analysis. In UCI, Cited by: §2.
  • [14] X. Huang and S. Belongie (2017) Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, Cited by: §3.2.
  • [15] J. Jia, Q. Ruan, and T. M. Hospedales (2019) Frustratingly easy person re-identification: generalizing person re-id in practice. In BMVC, Cited by: §1, §2, §3.2.
  • [16] X. Jin, C. Lan, W. Zeng, Z. Chen, and L. Zhang (2020) Style normalization and restitution for generalizable person re-identification. In CVPR, Cited by: §1, §2, §3.2, §4.1, Table 2.
  • [17] X. Jin, C. Lan, W. Zeng, and Z. Chen (2020) Feature alignment and restoration for domain generalization and adaptation. arXiv. Cited by: §2.
  • [18] D. P. Kingma and J. Ba (2015) Adam: a method for stochastic optimization. In ICLR, Cited by: §4.
  • [19] D. Li, Y. Yang, Y. Song, and T. M. Hospedales (2018) Learning to generalize: meta-learning for domain generalization. In AAAI, Cited by: Table 2.
  • [20] H. Li, S. J. Pan, S. Wang, and A. C. Kot (2018) Domain generalization with adversarial feature learning. In CVPR, Cited by: §2, §2.
  • [21] H. Li, Y. Wang, R. Wan, S. Wang, T. Li, and A. Kot (2020) Domain generalization for medical imaging classification with linear-dependency regularization. NeurIPS. Cited by: §2.
  • [22] P. Li, D. Li, W. Li, S. Gong, Y. Fu, and T. M. Hospedales (2021) A simple feature augmentation for domain generalization. In ICCV, Cited by: §2, §3.2.
  • [23] W. Li and X. Wang (2013) Locally aligned feature transforms across views. In CVPR, Cited by: §4, Table 1.
  • [24] W. Li, R. Zhao, T. Xiao, and X. Wang (2014)

    Deepreid: deep filter pairing neural network for person re-identification

    In CVPR, Cited by: §4, Table 1.
  • [25] W. Li, X. Zhu, and S. Gong (2018) Harmonious attention network for person re-identification. In CVPR, Cited by: §1.
  • [26] X. Li, Y. Dai, Y. Ge, J. Liu, Y. Shan, and L. Duan (2022) Uncertainty modeling for out-of-distribution generalization. ICLR. Cited by: §2, §3.2.
  • [27] S. Liao and L. Shao (2020) Interpretable and generalizable person re-identification with query-adaptive convolution and temporal lifting. In ECCV, Cited by: Table 3.
  • [28] C. C. Loy, T. Xiang, and S. Gong (2010) Time-delayed correlation analysis for multi-camera activity understanding. IJCV. Cited by: §4, Table 1.
  • [29] C. Luo, C. Song, and Z. Zhang (2020) Generalizing person re-identification by camera-aware invariance learning and cross-domain mixup. In ECCV, Cited by: §1.
  • [30] D. Mahajan, S. Tople, and A. Sharma (2021) Domain generalization using causal matching. In ICML, Cited by: §1.
  • [31] K. Muandet, D. Balduzzi, and B. Schölkopf (2013) Domain generalization via invariant feature representation. In ICML, Cited by: §2.
  • [32] A. Nichol, J. Achiam, and J. Schulman (2018) On first-order meta-learning algorithms. arXiv. Cited by: Table 2.
  • [33] A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer (2017) Automatic differentiation in pytorch. In NIPS-W, Cited by: §4.
  • [34] S. Qiao, C. Liu, W. Shen, and A. L. Yuille (2018) Few-shot image recognition by predicting parameters from activations. In CVPR, Cited by: Table 2.
  • [35] S. Shankar, V. Piratla, S. Chakrabarti, S. Chaudhuri, P. Jyothi, and S. Sarawagi (2018) Generalizing across domains via cross-gradient training. arXiv. Cited by: Table 2.
  • [36] J. Song, Y. Yang, Y. Song, T. Xiang, and T. M. Hospedales (2019) Generalizable person re-identification by domain-invariant mapping network. In CVPR, Cited by: §4.1, Table 2.
  • [37] Y. Sun, L. Zheng, Y. Li, Y. Yang, Q. Tian, and S. Wang (2019) Learning part-based convolutional features for person re-identification. IEEE TPAMI 43 (3). Cited by: Table 2.
  • [38] L. Wei, S. Zhang, W. Gao, and Q. Tian (2018) Person transfer gan to bridge domain gap for person re-identification. In CVPR, Cited by: §1, §4, Table 1.
  • [39] T. Xiao, S. Li, B. Wang, L. Lin, and X. Wang (2016) End-to-end deep learning for person search. arXiv. Cited by: §4, Table 1.
  • [40] F. Yang, Y. Cheng, Z. Shiau, and Y. F. Wang (2021) Adversarial teacher-student representation learning for domain generalization. In NeurIPS, Cited by: §2.
  • [41] S. Yu, F. Zhu, D. Chen, R. Zhao, H. Chen, S. Tang, J. Zhu, and Y. Qiao (2021) Multiple domain experts collaborative learning: multi-source domain generalization for person re-identification. arXiv. Cited by: §1, §2.
  • [42] P. Zhang, Q. Liu, D. Zhou, T. Xu, and X. He (2018) On the discrimination-generalization tradeoff in gans. In ICLR, Cited by: §4.2.
  • [43] X. Zhang, H. Luo, X. Fan, W. Xiang, Y. Sun, Q. Xiao, W. Jiang, C. Zhang, and J. Sun (2017) Alignedreid: surpassing human-level performance in person re-identification. arXiv. Cited by: Table 2.
  • [44] Y. Zhang, H. Zhang, Z. Zhang, D. Li, Z. Jia, L. Wang, and T. Tan (2021) Learning domain invariant representations for generalizable person re-identification. arXiv. Cited by: §1, §2.
  • [45] Z. Zhang, C. Lan, W. Zeng, X. Jin, and Z. Chen (2020) Relation-aware global attention for person re-identification. In CVPR, Cited by: §1.
  • [46] Y. Zhao, Z. Zhong, F. Yang, Z. Luo, Y. Lin, S. Li, and N. Sebe (2021) Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification. In CVPR, Cited by: §2, §3.2, §3.4, §4.1, Table 3.
  • [47] L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian (2015) Scalable person re-identification: a benchmark. In ICCV, Cited by: §4, Table 1.
  • [48] W. Zheng, S. Gong, and T. Xiang (2009) Associating groups of people.. In BMVC, Cited by: §4, Table 1.
  • [49] Z. Zheng, X. Yang, Z. Yu, L. Zheng, Y. Yang, and J. Kautz (2019) Joint discriminative and generative learning for person re-identification. In CVPR, Cited by: §1.
  • [50] Z. Zheng, L. Zheng, and Y. Yang (2017) Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In ICCV, Cited by: §4, Table 1.
  • [51] Z. Zhong, L. Zheng, Z. Luo, S. Li, and Y. Yang (2019) Invariance matters: exemplar memory for domain adaptive person re-identification. In CVPR, Cited by: §3.4.
  • [52] K. Zhou, Z. Liu, Y. Qiao, T. Xiang, and C. C. Loy (2021) Domain generalization in vision: a survey. arXiv. Cited by: §1.
  • [53] K. Zhou, Y. Yang, T. Hospedales, and T. Xiang (2020) Learning to generate novel domains for domain generalization. In ECCV, Cited by: §1, §1.