PaperlayerView exact revision on arXiv

Source-linked AI summary

Style Normalization and Restitution for Generalizable Person Re-identification

Xin Jin, Cuiling Lan, Wenjun Zeng, Zhibo Chen, Li Zhang

arXiv:2005.11037v1cs.CV

TL;DR

Person ReID models often generalize poorly across domains because camera and environment styles differ, while target-domain annotation or adaptation can be costly. This paper proposes SNR, which uses Instance Normalization to filter style variation, restores identity-relevant information discarded by normalization, and applies a dual causality loss for disentanglement. Across multiple benchmarks and settings, SNR achieves the best reported performance for both domain generalization and unsupervised domain adaptation ReID, with additional effectiveness shown for RGB-Infrared ReID and another backbone.

  • Problem

    Domain gaps cause fully supervised ReID methods to generalize poorly to unseen datasets, while target-domain annotation or adaptation adds practical cost.

  • Method

    SNR uses Instance Normalization to filter style variation, restitutes discarded identity-relevant features, and applies a dual causality loss to disentangle relevant and irrelevant features.

  • Results

    SNR achieves the best reported performance on both domain generalization and unsupervised domain adaptation ReID across several benchmarks and settings.

  • Takeaways & Limitations

    SNR is a plug-and-play module that enhances generalization capabilities in existing ReID architectures and also shows effectiveness for RGB-Infrared ReID and another backbone.

  • Takeaways & Limitations

    The framework assumes access to one or several annotated source datasets and direct deployment to unseen domains or datasets.

Abstract

from arXiv · show

Existing fully-supervised person re-identification (ReID) methods usually suffer from poor generalization capability caused by domain gaps. The key to solving this problem lies in filtering out identity-irrelevant interference and learning domain-invariant person representations. In this paper, we aim to design a generalizable person ReID framework which trains a model on source domains yet is able to generalize/perform well on target domains. To achieve this goal, we propose a simple yet effective Style Normalization and Restitution (SNR) module. Specifically, we filter out style variations (e.g., illumination, color contrast) by Instance Normalization (IN). However, such a process inevitably removes discriminative information. We propose to distill identity-relevant feature from the removed information and restitute it to the network to ensure high discrimination. For better disentanglement, we enforce a dual causal loss constraint in SNR to encourage the separation of identity-relevant features and identity-irrelevant features. Extensive experiments demonstrate the strong generalization capability of our framework. Our models empowered by the SNR modules significantly outperform the state-of-the-art domain generalization approaches on multiple widely-used person ReID benchmarks, and also show superiority on unsupervised domain adaptation.

1. Introduction

Person ReID must identify people across cameras, environments, and locations despite domain gaps, while retaining discrimination between similar-looking individuals. The proposed SNR framework normalizes style variations, restores identity-relevant information lost by normalization, and uses dual causality constraints for disentanglement.

  • Person ReID matches or identifies specific people across cameras, times, and locations, supporting many applications.
  • Style discrepancies across cameras and environments cause poor generalization to unseen datasets, including differences in illumination, hue, contrast, saturation, and resolution.
  • Robust ReID requires both generalization across environments and discrimination between different people with similar appearances.
  • Domain generalization avoids target-domain data, labels, and model updates, but ReID remains challenging because unseen target domains typically contain different identities.
  • SNR applies Instance Normalization to reduce style variation, then restitutes identity-relevant information from the residual between original and normalized features.
  • The dual causality loss encourages identity-relevant features to become more discriminative and identity-irrelevant features to become less discriminative after restitution or compensation.
  • SNR is presented as a plug-and-play module that improves generalization on unseen domains and significantly outperforms state-of-the-art domain-generalizable approaches across multiple benchmarks and settings.

2. Related Work

Prior ReID methods often generalize poorly across domains because of style discrepancies, while general domain-generalization methods typically assume shared label spaces that ReID does not have. The paper situates SNR as a style-normalization and restitution approach compatible with existing ReID networks.

  • Fully supervised ReID methods perform well on source-dataset tests but generalize poorly to unseen domains with different styles.
  • The SNR design uses Instance Normalization to reduce style discrepancies, then restitutes identity-relevant information discarded by normalization.
  • General domain-generalization methods commonly assume identical source and target label spaces, whereas ReID targets typically contain different identities and form an open-set problem.
  • Existing generalizable ReID work includes combining multiple source datasets, meta-learning domain-invariant models, and maintaining a shared memory bank for inconsistent label spaces.

3. Proposed Generalizable Person ReID

The framework targets direct deployment on unseen domains by combining style normalization with restitution of identity-relevant information. Its SNR module filters domain-specific variation, restores discriminative content from the residual, and uses dual causality constraints to separate relevant from irrelevant features.

  • Domain generalization trains on annotated source datasets and deploys directly to unseen domains without target data, labels, or model updates.
  • Style Normalization Phase: Instance Normalization reduces domain discrepancy by filtering instance-specific style information, but can also discard identity-relevant discriminative information.
  • Style Restitution Phase: SNR defines the residual as the difference between the original feature F and its style-normalized feature eF, then separates the residual into identity-relevant and identity-irrelevant parts.
  • Style Restitution Phase: A learned channel-attention vector adaptively distills identity-relevant residual channels for restitution, with an SE-like design based on pooling and fully connected layers.
  • Style Restitution Phase: The enhanced feature adds the distilled identity-relevant residual to the normalized feature, bringing same-identity samples closer and different-identity samples farther apart.
  • Dual Causality Loss Constraint: The dual causality loss encourages restitution of identity-relevant features to increase discrimination while compensation with identity-irrelevant features decreases it.

4. Experiments

Experiments evaluate SNR across generalizable ReID benchmarks, ablations, visualizations, and UDA extensions. SNR consistently improves generalization, feature robustness, and performance while adding little model-size overhead.

  • Experimental Setup: Experiments use Market1501, DukeMTMC-reID, CUHK03, MSMT17, PRID, GRID, VIPeR, and i-LIDS under single- and multi-source transfer settings.Performance is evaluated with Rank-1 CMC and mAP, using source→unseen-target configurations such as Market1501→Duke and Duke→Market1501.
  • Ablation Study: Baseline-SNR significantly outperforms baseline normalization schemes, including 8.5%, 6.7%, and 15.0% mAP gains over Baseline-IN on M→D, D→M, and D→GRID.The final scheme also outperforms all baseline schemes in generalization capability.
  • Complexity: The final SNR model has nearly the same size as the baseline, at 24.74 M versus 24.56 M parameters.This comparison is reported alongside the generalization experiments.
  • Design Choices of SNR: Adding SNR to all four ResNet-50 stages achieves the best performance, while stage-3 and stage-4 additions improve more than stage-1 and stage-2 additions.Content-adaptive and interaction-based disentanglement designs also improve performance, with gains of 3.9%/4.5% and 2.4%/2.9% mAP in the reported comparisons.
  • Visualization: SNR activation maps better focus on human-body regions and remain more consistent across contrast, illumination, and saturation changes than the baseline.Feature distributions show that IN eliminates the source-target domain gap, while restitution reduces the gap and makes same-identity samples less dispersive.
  • Comparison and Extension: SNR outperforms strong domain-generalization and UDA comparisons, including 6.9%/7.8% mAP over OSNet-IBN and 3.8%/3.4% over the second-best UDA method.It also improves Baseline+MAR by 22.9%/24.5% mAP and yields 8.4%, 8.2%, 11.0%, and 11.5% gains in four RGB-infrared settings.

5. Conclusion

The paper introduces SNR to separate style-related interference from identity-relevant information, improving generalizable person ReID. Experiments report strong performance across domain generalization, unsupervised domain adaptation, RGB-Infrared ReID, and another backbone.

  • SNR uses Instance Normalization to filter style variations and restitutes identity-relevant features discarded during normalization.A dual causality loss further disentangles identity-relevant and identity-irrelevant features.
  • SNR achieves the best reported performance on both domain generalization and unsupervised domain adaptation ReID settings.
  • SNR also improves performance on RGB-Infrared ReID and when used with another backbone.

1. Implementation Details

The experiments use ResNet-50-based models with SNR inserted across the first four convolutional stages, standard augmentation, and specified domain-generalization and adaptation training procedures.

  • Network Details: ResNet-50 serves as the backbone, with SNR modules added after the last layer of each of the first four convolutional stages.The input resolution is 256×128, and the final spatial downsampling operation is removed.
  • Data Augmentation: Training uses random cropping, horizontal flipping, label smoothing, color jittering, and disabled random erasing to support generalization.The passages state that random erasing can hurt cross-domain ReID because masking training-image regions encourages source-specific learning.
  • Training Details for Domain Generalization: Domain-generalization batches sample P = 24 identities and K = 4 images per identity, producing batch size B = P × K = 96.The batch construction follows identity-balanced sampling.
  • Optimization: Optimization uses Adam with a 20-epoch linear warm-up from 8×10−6, weight decay 5×10−4, and learning-rate halving every 40 epochs.The SNR model converges after 280 epochs and is trained on a single 32G NVIDIA-V100 GPU using PyTorch.
  • Training Details for Domain Adaptation: For adaptation, labeled source and unlabeled target batches are jointly used, with ReID and dual causality losses on source data and MAR pseudo-label mining on target data.Each adaptation batch contains 48 labeled source and 48 unlabeled target samples.

2. Details of Datasets

The paper evaluates generalizable ReID across commonly used large datasets and small cross-domain benchmarks, emphasizing substantial style discrepancies between datasets and within MSMT17.

  • Dataset Overview: Table 4 provides detailed information about the person ReID datasets used in the experiments.The table is the source for dataset-specific details referenced in this section.
  • Dataset Roles: Market1501, DukeMTMC-reID, CUHK03, and MSMT17 support supervised and unsupervised adaptation evaluations, while VIPeR, PRID2011, GRID, and i-LIDS assess cross-domain generalization.The four small datasets are not involved in training.
  • Evaluation Organization: Table 5 organizes performance comparisons by source dataset rows and target-dataset testing columns, including total source training-image counts.Results on already-seen testing domains are masked for supervised ReID.
  • Visual Dataset Comparison: Images from Market-1501, DukeMTMC-reID, CUHK03, MSMT17, PRID, GRID, VIPeR, and i-LIDS are resized to 256×128 for visual comparison.Figure 5 samples 10 identities from each ReID dataset.
  • Style Variation: The datasets exhibit style discrepancies, especially PRID and GRID, while MSMT17 also contains substantial style variation within the same dataset.These observations motivate examining style robustness across domains.

3. More Ablation Study Results

The ablation results indicate that SNR improves generalization by restituting identity-relevant information, while cross-domain performance also benefits from using more source datasets.

  • Ablation Observations: IN-related baselines improve generalization ability but reduce performance when training and testing occur in the same domain.This establishes a trade-off between normalization-based generalization and same-domain performance.
  • SNR Effectiveness: Baseline-SNR achieves superior generalization capability, attributed to restitution of identity-relevant information by the SNR modules.The comparison is presented in Table 5.
  • Source-Domain Scaling: Generalization performance on unseen target domains increases consistently as the number of source datasets increases.This observation concerns the multi-source setting reported in the ablation comparisons.
  • Activation Analysis: The activation maps of the proposed scheme are more consistent and invariant to style variants than those of the strong baseline.Figure 6 compares the baseline maps at the top with the proposed method's maps at the bottom.
  • Data Volume versus Generalization: MSMT17 contains 126k images, compared with about 33K images for Market1501 or Duke, yet its Baseline reaches only 9.8% mAP on GRID.On GRID, this is below Market1501-trained Baseline at 25.8% and Duke-trained Baseline at 14.5%; on VIPeR and iLIDS, MSMT17 is 3.8% to 12.5% higher in mAP than those sources.

4. More Visualization Analysis

The analysis examines how SNR affects feature divergence and identity separability across domains. It reports reduced cross-domain divergence and more compact, better-separated identity features on unseen Duke.

  • Feature Divergence Analysis: The analysis compares Baseline, Baseline-IN, and SNR using Market1501-to-Duke feature distributions and cross-domain divergence.The evaluation samples 500 images from each dataset for feature-divergence analysis.
  • Feature Divergence Analysis: SNR reduces feature divergence across convolutional stages compared with Baseline and Baseline-IN.Feature divergence is measured using symmetric KL divergence between Market1501 and Duke features.
  • Feature Divergence Analysis: Stage-4 feature divergence exceeds Stage-3 divergence because high-level identity-discriminative semantics may increase domain divergence.The passage attributes this pattern to the stronger identity-related content of Stage-4 features.
  • Final Feature Visualization: On unseen Duke, SNR produces more compact same-identity feature clusters and pushes different identities farther apart than Baseline.The comparison uses t-SNE visualizations of final ReID feature vectors, with colors denoting identities.

5. Comparison with State-of-the-Arts (Complete version)

The complete comparison evaluates SNR against recent generalizable ReID methods under matched training conditions. SNR substantially outperforms both the Baseline and DIMN across several target benchmarks.

  • State-of-the-Art Comparison: 14.6%/6.6%/1.2%/11.5% mAP gains over DIMN occur on PRID/GRID/VIPeR/i-LIDS, respectively.Both methods are trained on the same five datasets under the same experimental setting.
  • State-of-the-Art Comparison: SNR outperforms the Baseline by up to 22.7% in mAP on PRID.The comparison is reported for the same five-source-dataset training setting.

6. Performance on Another Backbone

SNR transfers to another backbone and to RGB-Infrared person ReID. It improves OSNet cross-domain performance and achieves state-of-the-art results on SYSU-MM01.

  • OSNet Backbone: OSNet-SNR exceeds OSNet-IBN by 5.0% and 5.5% in mAP for M→D and D→M, respectively.SNR modules are inserted between OS-Blocks using OSNet’s official released weights and codes.
  • RGB-Infrared Cross-Modality ReID: SNR is evaluated on RGB-Infrared cross-modality ReID, where RGB and Infrared images exhibit a large style discrepancy.Evaluation uses the SYSU-MM01 benchmark and averages ten repeated random gallery-probe splits.
  • RGB-Infrared Cross-Modality ReID: 11.5% mAP is the largest reported SNR gain over Baseline across the four SYSU-MM01 experimental settings.The four gains are 8.4%, 8.2%, 11.0%, and 11.5%, and the resulting scheme achieves state-of-the-art performance.
Loading 2005.11037v1…