Source-linked AI summary
Meta Batch-Instance Normalization for Generalizable Person Re-Identification
Seokeon Choi, Taekyung Kim, Minki Jeong, Hyoungseob Park, Changick Kim
TL;DR
DG Re-ID seeks strong performance on unseen domains without target updates, yet existing normalization can discard identity-discriminative information. MetaBIN combines learnable batch-instance normalization with meta-learning that simulates normalization failures, and it reports state-of-the-art results on domain-generalization and cross-domain Re-ID benchmarks.
Problem
Existing normalization-based DG Re-ID methods can lose discriminative information, while learnable BN-IN balancing can overfit source-domain styles and generalize poorly.
Method
MetaBIN applies meta-learning to learnable BN-IN balancing parameters by simulating under-style- and over-style-normalization scenarios in alternating update episodes.
Results
MetaBIN achieves state-of-the-art performance on a large-scale domain-generalization Re-ID benchmark and the cross-domain Re-ID problem.
Takeaways & Limitations
MetaBIN improves generalization to unseen domains while avoiding additional data augmentation or complicated network design.
Takeaways & Limitations
DG Re-ID uses heterogeneous domains with disjoint identity label spaces, unlike homogeneous DG settings.
Abstract
from arXiv · showhide
Although supervised person re-identification (Re-ID) methods have shown impressive performance, they suffer from a poor generalization capability on unseen domains. Therefore, generalizable Re-ID has recently attracted growing attention. Many existing methods have employed an instance normalization technique to reduce style variations, but the loss of discriminative information could not be avoided. In this paper, we propose a novel generalizable Re-ID framework, named Meta Batch-Instance Normalization (MetaBIN). Our main idea is to generalize normalization layers by simulating unsuccessful generalization scenarios beforehand in the meta-learning pipeline. To this end, we combine learnable batch-instance normalization layers with meta-learning and investigate the challenging cases caused by both batch and instance normalization layers. Moreover, we diversify the virtual simulations via our meta-train loss accompanied by a cyclic inner-updating manner to boost generalization capability. After all, the MetaBIN framework prevents our model from overfitting to the given source styles and improves the generalization capability to unseen domains without additional data augmentation or complicated network design. Extensive experimental results show that our model outperforms the state-of-the-art methods on the large-scale domain generalization Re-ID benchmark and the cross-domain Re-ID problem. The source code is available at: https://github.com/bismex/MetaBIN.
1. Introduction
Generalizable Re-ID targets unseen domains without target-domain updates, but normalization must balance removing style variation against retaining identity-discriminative information. MetaBIN addresses this by simulating normalization failures in meta-learning and reports state-of-the-art performance across domain-generalization and cross-domain Re-ID.
- Supervised Re-ID performs well within separated dataset splits but degrades on unseen target domains and requires costly labeling.
- Domain generalization avoids target images, but existing meta-learning methods largely assume homogeneous label spaces and classification tasks rather than Re-ID.
- BN can fail to distinguish identities under unexpected unseen styles, whereas IN can remove identity-discriminative information.
- MetaBIN generalizes batch-instance normalization by simulating under-style- and over-style-normalization scenarios within a meta-learning pipeline.
- Learnable balancing parameters combine BN and IN, while meta-train loss and cyclic inner-updating diversify virtual simulations to improve generalization.
- MetaBIN achieves state-of-the-art performance on a large-scale domain-generalization Re-ID benchmark and the cross-domain Re-ID problem.
2. Related Work
Related work combines normalization and meta-learning to improve cross-domain robustness, but existing approaches leave important limitations for DG Re-ID. MetaBIN combines these directions while avoiding additional test modules and data augmentation.
- DG Re-ID methods seek robust performance on unseen target domains without additional updates, using either single or multiple source datasets.
- Existing BN-IN methods reduce style information, but IN can inevitably remove discriminative information and non-parametric methods depend on layer placement.
- MetaBIN combines normalization layers with meta-learning to generalize DG Re-ID without an additional test module or data augmentation.
- MAML-based domain-generalization methods simulate train-test domain shifts, but their homogeneous-DG design limits direct application to DG Re-ID.
3. Proposed Method
MetaBIN formulates DG Re-ID with heterogeneous source identities and learns normalization layers that combine BN and IN while simulating challenging unseen-style cases through meta-learning.
- Problem formulation: DG Re-ID trains on aggregated labeled source domains and retrieves identities in unseen target domains without additional model updating.
- Batch-instance normalization: The batch-instance normalization layer combines BN and IN using channel-wise learnable balancing parameters and separate affine transformations.
- Batch-instance normalization: Learnable BN–IN balancing can overfit source-domain styles, causing unsuccessful normalization on unseen target domains.
- Meta Batch-Instance Normalization: MetaBIN separates base-model updates from balancing-parameter updates, applying meta-learning only to the balancing parameters.
- Meta Batch-Instance Normalization: Randomly split source domains into meta-train and meta-test sets to mimic train–test domain shifts during alternating inner and meta updates.
- Meta-train simulations: Meta-train losses simulate over-style-normalization and under-style-normalization by biasing balancing parameters toward IN and BN, respectively.
4. Experiments
Experiments evaluate MetaBIN on large-scale and cross-domain Re-ID benchmarks, then analyze its components, sampling strategy, normalization behavior, balancing parameters, and meta-train losses. Across these evaluations, MetaBIN improves generalization while avoiding additional memory usage, data augmentation, or complicated network design.
- Large-scale DG Re-ID benchmark: MetaBIN outperformed all competing methods by a significant margin in average performance on the large-scale DG Re-ID benchmark.The evaluation used both MobileNetV2 and ResNet-50 backbones across multiple unseen target datasets.
- Cross-domain Re-ID: MetaBIN achieved state-of-the-art performance on the cross-domain Re-ID problem using unseen target datasets from single labeled source domains.The cross-domain setting evaluates generalization from one source dataset to an unseen target dataset.
- Influence of model components: Updating balancing parameters in meta-learning substantially improves generalization by alleviating overfitting to the given source domains.The ablation also reports that triplet loss is more suitable than cross-entropy for simulating train-test domain shifts, while the meta-train losses are complementary.
- Influence of domain-level sampling: Using 3 meta-train domains and 2 meta-test domains produced the highest performance and lowest memory usage, while the pipeline stayed within the 7,895MiB base-model memory budget.Separating meta-learning and base-model updates avoids increasing memory usage and provides an advantage over MLDG's full-parameter meta-updating.
- Analysis from the perspective of normalization: MetaBIN surpassed other normalization methods in DG experiments and shortened query-gallery distances in t-SNE analysis, while also improving supervised-setting performance.The analysis attributes these gains to simulating unsuccessful generalization scenarios during meta-learning.
- Analysis on the balancing parameters: Most balancing parameters retain BN properties, while selected channels shift toward IN; this combination removes some instance-specific style information and promotes generalization.The analysis specifically reports that shifting middle- and shallow-layer channels toward IN alleviates overfitting.
5. Conclusion
MetaBIN is a generalizable person re-identification framework that addresses overfitting by simulating unsuccessful generalization scenarios. Its meta-train loss and cyclic inner-updating diversify simulations and improve performance across domain-generalization settings.
- MetaBIN alleviates overfitting by investigating unsuccessful generalization scenarios.
- Its meta-train loss and cyclic inner-updating method diversify virtual simulations for unseen domains.
- Experiments in single- and multi-source domain generalization demonstrate superiority over state-of-the-art methods.
1. More Analysis on Hyperparameters
The hyperparameter analyses examine loss weighting, cyclic inner-updates, and meta-optimization step size. All loss components contribute, five-epoch cycles perform best, and larger final step sizes shift balancing toward instance normalization.
- Meta-train loss: Each meta-train loss weight performs best at 1.0, while removing any component deteriorates performance.The components are L_scat, L_shuf, and L_tr.
- Cyclic inner-updates: Five epochs produce the highest performance for cyclic inner-updates, so the cycle period is set to five epochs.This corresponds to 9,245 iterations, with β oscillating every five epochs.
- Meta-optimization step size: As γ increases, balancing parameters become more biased toward instance normalization.The appropriate γ depends on style variation between domains.
- Training analysis: The meta-train losses are tracked during training to analyze optimization behavior.
2. More Analysis on Balancing Parameters
Balancing-parameter analyses show that MetaBIN distributes normalization across layers differently from BIN and adapts this distribution to architecture, dataset, and training stage. Camera-domain variation within single-source datasets produces little bias toward instance normalization.
- Network and dataset effects: MetaBIN normalizes styles across overall layers except the last, regardless of network architecture.BIN instead applies instance-style normalization only in shallow and deep layers.
- Network and dataset effects: Balancing parameters are hardly biased toward instance normalization on Market-1501 and DukeMTMC-ReID.Camera-domain discrepancy within each single-source dataset is smaller than discrepancy across multiple-source domains.
- Training dynamics: At training onset, layers 1, 2, 8, and 9 sharply reduce their mean balancing parameters before rebounding.The initial reduction removes useless style information to improve discrimination rather than generalization capability.
- Architecture comparison: Table 6 compares performance, memory usage, and inference time across network structures.Inference time is measured per image with mini-batch size 64.
3. Various architecture designs
The study evaluates domain-generalizable person re-identification across varied network architectures, including MobileNetV2 and ResNet variants, for practical use without target-domain updates.
- The experiments cover MobileNetV2 and ResNet variants to assess practical domain-generalizable Re-ID across architectures.The task targets good unseen-domain performance without additional target-image updates.