Research map / Label-efficient and robust learning / Domain adaptation and generalization
Generalizable Person Re-identification via Balancing Alignment and Uniformity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that data augmentation (Random Erasing, RandAugment) helps in-distribution but hurts out-of-distribution person re-ID, and proposes BAU, which balances alignment and uniformity losses on original and augmented images with a reliability weighting and a domain-specific uniformity loss.
Problem
Domain generalizable person re-identification suffers because data augmentations such as Random Erasing and RandAugment improve source-domain performance but degrade unseen-domain performance, and the reason for this polarized effect is unexplained.
Approach
Diagnoses the polarized augmentation effect as a loss of uniformity (sparse representation space) and regularizes training by applying alignment and uniformity losses to both original and augmented images, weighting alignment pairs by k-reciprocal-neighbor Jaccard similarity, plus a memory-bank uniformity loss computed within each source domain.
Key results
- On Protocol-3, BAU reaches 52.3% average mAP and 65.7% average Rank-1 across three leave-one-out targets, versus 46.1%/59.0% for the no-augmentation baseline in the ablation.
- On Protocol-1, BAU averages 77.2% mAP and 69.5% Rank-1 over PRID, GRID, VIPeR and iLIDs, above the prior best reported averages.
- On Protocol-2 BAU improves average mAP to 48.1% from 45.3% for ACL, and on Protocol-3 to 52.3% from 49.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BAU instead of ACL | DG re-ID · Protocol-2 average | mAP: 45.3 -> 48.1 | Helped |
| Use BAU instead of ACL | DG re-ID · Protocol-3 average | mAP: 49.3 -> 52.3 | Helped |
| Apply cross-entropy loss to augmented images instead of training without augmented images (ablation) | BAU training · Protocol-3 average | mAP: 42.1 -> 37.7 | Hurt |
| Add alignment loss for augmented images (ablation) | BAU training · Protocol-3 average | mAP: 42.1 -> 46.2 | Helped |
| Add uniformity loss to alignment loss (ablation) | BAU training · Protocol-3 average | mAP: 46.2 -> 50.1 | Helped |
| Add domain-specific uniformity loss (ablation) | BAU training · Protocol-3 average | mAP: 50.1 -> 52.3 | Helped |
| Add alignment weighting (ablation) | BAU alignment loss · Protocol-3 average | mAP: 48.9 -> 50.0 | Helped |
| Use domain-specific prototypes instead of prototypes from the entire domain (ablation) | BAU domain-specific uniformity loss · Protocol-3 average | mAP: 50.0 -> 52.3 | Helped |
| Use all three augmentations instead of no augmentations (ablation) | BAU · Protocol-3 average | mAP: 46.1 -> 52.3 | Helped |
| Increase augmentation probability from 0.5 to 0.75 or 1.0 (ablation) | BAU · Protocol-3 average | mAP: 52.3 -> 51.6; 52.3 -> 49.3 | Hurt |
| Apply BAU rather than the baseline without BAU | MobileNetV2 · Protocol-2 average | mAP: 26.1 -> 33.2 | Helped |
| Apply BAU rather than the baseline without BAU | ViT-B/16 · Protocol-2 average | mAP: 33.2 -> 40.3 | Helped |
Surprising findings
- Random Erasing and RandAugment, though known to hurt cross-domain re-ID, become beneficial when combined with the proposed alignment and uniformity regularization.
- Even with no augmentation (p=0), the BAU regularization outperforms the baseline.
Limitations
- May struggle under very large domain shifts, as acknowledged by the authors.
- Uses only image-level augmentations; adversarial and feature-level augmentation are not explored.
Experimental setup
Datasets: Market-1501 MSMT17 CUHK02 CUHK03 CUHK-SYSU PRID GRID VIPeR iLIDs
Benchmarks: Market-1501 MSMT17 CUHK03 PRID GRID VIPeR iLIDs
Metrics: mAP Rank-1
Baselines: SNR QAConv M3L MetaBIN META ACL GDNorm DDAN StyCon DualNorm
Builds on: alignment and uniformity (Wang and Isola) RandAugment Random Erasing batch-hard triplet loss momentum prototype update
Closest papers by meaning
- Dual Distribution Alignment Network for Generalizable Person Re-Identification (AAAI 2021 · similarity 0.91)
- Meta Distribution Alignment for Generalizable Person Re-Identification (CVPR 2022 · similarity 0.91)
- Generalizable Person Re-identification via Self-Supervised Batch Norm Test-Time Adaption (AAAI 2022 · similarity 0.89)
- Generalizable Person Re-Identification by Domain-Invariant Mapping Network (CVPR 2019 · similarity 0.89)
- Unsupervised Domain Adaptation for Person Re-identification via Heterogeneous Graph Alignment (AAAI 2021 · similarity 0.89)
- Diversity-Authenticity Co-constrained Stylization for Federated Domain Generalization in Person Re-identification (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.