[Paper Review] Debiased Batch Normalization via Gaussian Process for Generalizable Person Re-Identification

Jiawei Liu, Zhipeng Huang+3 more|arXiv (Cornell University)|Mar 3, 2022

Video Surveillance and Tracking Methods4 citations

TL;DR

This paper proposes GDNorm, a novel debiased batch normalization method for generalizable person re-identification that models domain-specific batch normalization statistics as a dynamically self-refining Gaussian process to reduce bias toward unseen domains. By leveraging the mean path of the Gaussian process for inference, GDNorm improves generalization with minimal computational cost, achieving state-of-the-art performance with a 2.8% R1 gain and 4.7% mAP improvement over baselines on unseen domains.

ABSTRACT

Generalizable person re-identification aims to learn a model with only several labeled source domains that can perform well on unseen domains. Without access to the unseen domain, the feature statistics of the batch normalization (BN) layer learned from a limited number of source domains is doubtlessly biased for unseen domain. This would mislead the feature representation learning for unseen domain and deteriorate the generalizaiton ability of the model. In this paper, we propose a novel Debiased Batch Normalization via Gaussian Process approach (GDNorm) for generalizable person re-identification, which models the feature statistic estimation from BN layers as a dynamically self-refining Gaussian process to alleviate the bias to unseen domain for improving the generalization. Specifically, we establish a lightweight model with multiple set of domain-specific BN layers to capture the discriminability of individual source domain, and learn the corresponding parameters of the domain-specific BN layers. These parameters of different source domains are employed to deduce a Gaussian process. We randomly sample several paths from this Gaussian process served as the BN estimations of potential new domains outside of existing source domains, which can further optimize these learned parameters from source domains, and estimate more accurate Gaussian process by them in return, tending to real data distribution. Even without a large number of source domains, GDNorm can still provide debiased BN estimation by using the mean path of the Gaussian process, while maintaining low computational cost during testing. Extensive experiments demonstrate that our GDNorm effectively improves the generalization ability of the model on unseen domain.

Motivation & Objective

• Address the challenge of domain shift in person re-identification when testing on unseen domains without access to target data.
• Overcome the bias introduced by limited source domains in batch normalization statistics, which hinders model generalization.
• Improve generalization performance by modeling domain-specific BN parameters as a self-refining Gaussian process that approximates the true data distribution.
• Reduce computational cost compared to ensemble methods while maintaining or improving performance on unseen domains.
• Enable effective feature representation learning for open-set retrieval in person re-identification under domain generalization settings.

Proposed method

• Train multiple domain-specific batch normalization layers on each source domain to capture unique domain characteristics.
• Use the parameters of these domain-specific BN layers as observations to learn a Gaussian process that models the underlying distribution of domain statistics.
• Dynamically self-refine the Gaussian process by sampling paths from it, using them to re-optimize the source domain BN parameters in a feedback loop.
• Estimate the optimal BN statistics for unseen domains using the mean path of the Gaussian process, which represents the most likely distribution of unseen domains.
• Apply the mean path as the BN layer parameters during inference, avoiding the need to train or deploy multiple models.
• Control the self-refinement process via a sampling ratio λ in the loss function to balance domain-specific and global statistics.

Experimental results

Research questions

• RQ1 Can modeling batch normalization statistics as a Gaussian process reduce bias in feature representation for unseen domains in person re-identification?
• RQ2 How does the self-refining mechanism of the Gaussian process improve the accuracy of BN statistics estimation for unseen domains?
• RQ3 Does using the mean path of the Gaussian process as the BN estimator outperform individual domain-specific models or ensemble methods in generalization performance?
• RQ4 What is the optimal trade-off between domain-specific and global statistics in the self-refinement process, as controlled by the sampling ratio λ?
• RQ5 Can GDNorm achieve state-of-the-art performance with low inference cost, even with only a few source domains?

Key findings

• GDNorm improves R1 accuracy by 2.8% and mAP by 4.7% compared to baseline methods on unseen domains under Protocol-2.
• The mean path of the Gaussian process achieves the best average performance, outperforming individual domain-specific models (e.g., D, MT, C3) and sampled paths.
• The optimal sampling ratio λ is 0.6, where the model achieves 86.5% R1 and 68.2% mAP, with performance degrading when λ exceeds 0.6.
• Inference with GDNorm takes only 0.09 seconds per batch, which is one-third the time of an ensemble model with three domain-specific networks.
• The self-refining mechanism significantly improves the accuracy of the Gaussian process estimation by iteratively refining BN parameters using sampled paths.
• The method maintains strong generalization even with a small number of source domains, demonstrating robustness to limited data.