Research map / Label-efficient and robust learning / Domain adaptation and generalization
Diversity-Authenticity Co-constrained Stylization for Federated Domain Generalization in Person Re-identification
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.
Proposes DACS, a federated domain generalization method for person re-ID that generates diverse and authentic stylized local data via a per-domain learnable style transformation model constrained by Wasserstein diversity and entropy-based authenticity losses.
Problem
Federated domain generalization for person re-identification (FedDG re-ID), where source domains are decentralized and cannot be pooled, so standard stylization-based DG methods cannot ensure diverse or authentic generated data.
Approach
A per-client style transformation model (learnable channel/spatial mean and std) is trained with a diversity loss that enlarges the simplified 2-Wasserstein distance between original and stylized feature statistics, and an authenticity loss that uses entropy from the local-side global model (should be low-uncertainty) and local model (should be high-uncertainty) to keep stylized samples realistic.
Key results
- On MS+C2+C3→M with ResNet-50, DACS reaches 36.3% mAP and 61.2% rank-1, the best among compared federated methods (SNR 32.7/59.4, CrossStyle 32.5/59.6).
- With ViT backbone on MS+C2+C3→M, Ours (ViT) achieves 45.4% mAP and 70.7% rank-1, versus FedPav (ViT) 37.4/62.6 and CrossStyle (ViT) 41.4/65.8.
- Combining DACS with SNR improves MS+C2+C3→M to 37.7% mAP and 65.9% rank-1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DACS stylization (ablation) | FedDG re-ID with ResNet-50 · MS+C2+C3→M | mAP: 25.4 -> 36.3 | Helped |
| Add DACS stylization (ablation) | FedDG re-ID with ResNet-50 · MS+C2+M→C3 | mAP: 22.5 -> 30.7 | Helped |
| Replace unconstrained random STM styles with a diversity constraint (ablation) | FedDG re-ID · MS+C2+C3→M | mAP: 25.1 -> 31.9 | Helped |
| Replace unconstrained random STM styles with an authenticity constraint (ablation) | FedDG re-ID · MS+C2+C3→M | mAP: 25.1 -> 34.5 | Helped |
| Use both diversity and authenticity constraints instead of either alone (ablation) | FedDG re-ID · MS+C2+C3→M | mAP: 31.9 / 34.5 -> 36.3 | Helped |
| Use both diversity and authenticity constraints instead of either alone (ablation) | FedDG re-ID · MS+C2+M→C3 | mAP: 29.1 / 27.3 -> 30.7 | Helped |
| Use Wasserstein distance rather than JSD for the diversity metric (ablation) | DACS · MS+C2+C3→M | mAP: 34.2 -> 36.3 | Helped |
| Use Wasserstein distance rather than KL divergence for the diversity metric (ablation) | DACS · MS+C2+C3→M | mAP: 33.6 -> 36.3 | Helped |
| Transfer style at the spatial level rather than the channel level (ablation) | DACS · MS+C2+C3→M | mAP: 30.0 -> 36.3 | Helped |
| Use the complete authenticity constraint rather than only H(fG(x)) < H(fG(x′)) (ablation) | DACS · MS+C2+C3→M | mAP: 33.3 -> 36.3 | Helped |
| Use the complete authenticity constraint rather than only H(fG(x′)) < H(fL(x′)) (ablation) | DACS · MS+C2+C3→M | mAP: 34.6 -> 36.3 | Helped |
| Combine DACS with SNR | FedDG re-ID · MS+C2+C3→M | mAP: 32.7 -> 37.7 | Helped |
Surprising findings
- Random style transformation without constraints (RandStyle) does not outperform the vanilla federated baseline, showing unconstrained stylization is ineffective.
Limitations
- Authenticity is judged only via entropy from the two re-ID models, not by a dedicated realism measure.
- Early-training stylized images show unrealistic contrast and illumination, which the authors note may degrade performance.
- Evaluated only on four person re-ID benchmarks; DukeMTMC-reID withdrawn and excluded.
Experimental setup
Datasets: MSMT-17 CUHK02 CUHK03 Market-1501
Benchmarks: Market-1501 CUHK02 CUHK03 MSMT-17
Metrics: mAP rank-1
Baselines: MOON SCAFFOLD FedPav FedReID MixStyle CrossStyle SNR Joint
Builds on: FedAvg MixStyle FedPav SNR
Closest papers by meaning
- FedARKS: Federated Aggregation via Robust and Discriminative Knowledge Selection and Integration for Person Re-identification (AAAI 2026 · similarity 0.89)
- Domain-Aware Suppression and Aggregation for Federated DG ReID (AAAI 2026 · similarity 0.89)
- Meta Batch-Instance Normalization for Generalizable Person Re-Identification (CVPR 2021 · similarity 0.87)
- Style Normalization and Restitution for Generalizable Person Re-Identification (CVPR 2020 · similarity 0.87)
- Decentralised Learning from Independent Multi-Domain Labels for Person Re-Identification (AAAI 2021 · similarity 0.87)
- Generalizable Person Re-identification via Balancing Alignment and Uniformity (NeurIPS 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.