Source-linked AI summary
Performance Optimization for Federated Person Re-identification via Benchmark Analysis
Weiming Zhuang, Yonggang Wen, Xuesen Zhang, Xin Gan, Daiying Yin, Dongzhan Zhou, Shuai Zhang, Shuai Yi
TL;DR
Centralized person ReID data raises privacy concerns, and statistical heterogeneity remains little studied in FedReID. The paper constructs a benchmark and proposes knowledge distillation and dynamic weight adjustment. Its benchmark analysis identifies bottlenecks, while distillation significantly elevates CUHK03-NP performance.
Problem
Centralized ReID training raises privacy risks, and little work studies the non-IID data and unbalanced volumes that challenge FedReID.
Source: §1 · p. 1Method
The paper constructs FedReIDBench and proposes knowledge distillation and dynamic weight adjustment to address convergence and imbalance-related performance decay.
Source: §1 · p. 2Results
The benchmark analysis identifies FedReID bottlenecks, and knowledge distillation significantly elevates performance on CUHK03-NP.
Source: §27 · p. 7, §6 · p. 8Takeaways & Limitations
FedReIDBench provides a baseline for future research on federated person re-identification.
Source: §1 · p. 2Future work will consider system heterogeneity.
Source: §6 · p. 9
Abstract
from arXiv · showhide
Federated learning is a privacy-preserving machine learning technique that learns a shared model across decentralized clients. It can alleviate privacy concerns of personal re-identification, an important computer vision task. In this work, we implement federated learning to person re-identification (FedReID) and optimize its performance affected by statistical heterogeneity in the real-world scenario. We first construct a new benchmark to investigate the performance of FedReID. This benchmark consists of (1) nine datasets with different volumes sourced from different domains to simulate the heterogeneous situation in reality, (2) two federated scenarios, and (3) an enhanced federated algorithm for FedReID. The benchmark analysis shows that the client-edge-cloud architecture, represented by the federated-by-dataset scenario, has better performance than client-server architecture in FedReID. It also reveals the bottlenecks of FedReID under the real-world scenario, including poor performance of large datasets caused by unbalanced weights in model aggregation and challenges in convergence. Then we propose two optimization methods: (1) To address the unbalanced weight problem, we propose a new method to dynamically change the weights according to the scale of model changes in clients in each training round; (2) To facilitate convergence, we adopt knowledge distillation to refine the server model with knowledge generated from client models on a public dataset. Experiment results demonstrate that our strategies can achieve much better convergence with superior performance on all datasets. We believe that our work will inspire the community to further explore the implementation of federated learning on more computer vision tasks in real-world scenarios.
1 INTRODUCTION
Person ReID depends on centralized personal images, creating privacy risks, while little work examines how statistical heterogeneity affects federated ReID. This work builds a benchmark and proposes optimization methods for FedReID.
1 INTRODUCTION
- Centralizing the personal images needed for ReID creates potential privacy risks and has limited the task’s development.The introduction describes personal information in these images as the source of the privacy concern.Source: §1 · p. 1
- Little work studies statistical heterogeneity in FedReID, although non-IID data and unbalanced volumes are identified as key real-world challenges.Prior work cited in the introduction links non-IID data with reduced federated-learning performance and convergence challenges.Source: §1 · p. 2
- FedReIDBench combines nine varied ReID datasets, representative federated scenarios, a suitable algorithm, standardized evaluation, and a reference implementation.The dataset variation simulates real-world non-IID and unbalanced data.Source: §1 · p. 2
- The work proposes knowledge distillation for convergence and dynamic weight adjustment to address performance decay from imbalanced datasets.Knowledge distillation addresses challenges associated with non-IID data; dynamic weight adjustment targets imbalance in aggregation.Source: §1 · p. 2
2 RELATED WORK
Person ReID matches people across gallery images, but training requires centralizing personal data. Prior federated-learning work addresses other benchmarks and non-IID challenges; this paper brings these concerns to ReID.
2.1 Person Re-identification
- Person ReID retrieves gallery images matching a query identity, and training requires centralizing personal data that raises privacy risks.Deep neural networks are described as improving representative feature extraction over hand-crafted operators.Source: §2.1 · p. 2
2.2 Federated Learning
- Prior federated-learning benchmarks target image classification, language tasks, and object detection [2], while this work benchmarks ReID and adopts unlabelled-data distillation inspired by prior non-IID solutions [29] [27].The related-work discussion also describes FedProx as restricting local updates toward the global model.Source: §2.2 · pp. 2–3
3 FEDERATED PERSON REID BENCHMARK
FedReIDBench defines a benchmark using nine diverse datasets, two federated scenarios, a model structure, training algorithms, performance metrics, and reference implementations. Its design accounts for dataset heterogeneity and differing client classifiers.
3.1 Datasets
- FedReIDBench uses nine datasets with varying image counts, identities, scenes, and camera views to simulate statistical heterogeneity and domain gaps.The volume disparities model imbalanced data across edges, while domain gaps model non-IID data.Source: §3.1 · p. 3
3.2 Federated Scenarios
- Federated-by-camera makes each camera a client communicating directly with the server, while federated-by-dataset uses edge servers to collect data from multiple cameras.The former requires cameras to compute deep models; the latter represents a client-edge-cloud architecture.Source: §3.2 · p. 3, Figure 2 · p. 3
3.3 Model Structure
- The benchmark uses an IDE model with a ResNet-50 backbone, while client identity classifiers can differ with the number of identities.These classifier differences affect the federated algorithm because client model structures are not identical.Source: §3.3 · p. 3
3.4 Federated Learning Algorithms
- FedPav supports clients with partially different models by aggregating their shared backbone parameters while retaining their local identity classifiers.Each round sends the global model to selected clients, which train locally and upload updated backbone parameters for server aggregation.Source: §3.4 · p. 3, Figure 3 · p. 4
- FedPav's global and client-local models are compared with local training, and identity classifiers are omitted during evaluation.ReID evaluation uses query images to search for similar gallery images.Source: §3.4 · pp. 3–4
3.6 Reference Implementation
- FedReIDBench provides reference implementations, including FedPav and optimization methods, plus dataset-preprocessing scripts.Source: §3.6 · p. 4
4 BENCHMARK ANALYSIS
The benchmark compares federated-by-camera and federated-by-dataset settings, finding weaker performance in the former and exposing performance and convergence challenges in the latter. In the federated-by-dataset setting, smaller local epochs improve accuracy at higher communication cost, while aggregation can hurt large-dataset performance.
4.1 Federated-by-camera Scenario
- Federated-by-camera has the lowest accuracy on Market-1501 and CUHK03-NP, while federated-by-identity outperforms it on both datasets.The paper suggests cross-camera knowledge is essential because camera-specific clients do not generalize well to multi-camera evaluation.Source: Table 2 · p. 5, §4.1 · p. 5
4.2 Federated-by-dataset Scenario
- With batch size B=32, performance is best in most datasets when local epochs E=1 rather than using larger local-epoch counts.The experiments fix B=32 and 300 total rounds; the paper reports that E=1 greatly outperforms E=5 and E=10 across all datasets.Source: Figure 5 · p. 5, §4.2 · p. 5
- Smaller local-epoch counts improve performance but increase communication cost.The paper identifies local epochs as a trade-off between communication cost and performance.Source: §4.2 · p. 5
- FedPav outperforms local training on smaller datasets such as CUHK01 and VIPeR, but underperforms it on larger datasets such as MSMT17 and Market-1501.The authors suggest larger-dataset models dominate aggregation, while models from smaller datasets may gain more from knowledge from other clients.Source: §4.2 · p. 5, Figure 6 · p. 6
- Before aggregation, local models outperform the federated model on all datasets, and local training exceeds those local models on large datasets.The authors interpret this as evidence that server aggregation could better integrate client knowledge and that FedPav has bottlenecks.Source: §4.2 · pp. 5–6, Figure 6 · p. 6
- FedPav rank-1 accuracy fluctuates during 300 communication rounds on DukeMTMC-reID and CUHK03-NP, indicating convergence difficulty under the benchmark's non-IID data.Evaluation was computed every 10 rounds with E=1 and B=32.Source: §4.2 · p. 6
5 PERFORMANCE OPTIMIZATION
The optimization section applies knowledge distillation to FedPav to improve performance and convergence. It transfers knowledge from client models to the server using an unlabelled public dataset.
5 PERFORMANCE OPTIMIZATION
- The section investigates knowledge distillation, weight adjustment, and their combination as performance-optimization methods.Source: §5 · p. 6
5.1 Knowledge Distillation
- Knowledge distillation transfers knowledge from client models acting as teachers to the server model acting as student.The approach targets FedPav's convergence difficulty and the gap between local and federated model performance.Source: §5.1 · p. 6
- The method uses an unlabelled CUHK02 dataset as a public example dataset for generating soft labels from clients.CUHK02 contains 1,816 identities in 7,264 images.Source: §5.1 · p. 6
27 return
The paper adjusts FedPav’s aggregation weights using client-model changes and combines this adjustment with knowledge distillation (KD). The combination improves performance and convergence on the reported datasets, while KD alone chiefly helps convergence and its accuracy effect depends on the dataset.
27 return
- The KD procedure averages client-generated soft labels from a shared public dataset and uses them to fine-tune the server model, mitigating aggregation instability.The shared dataset identified in the procedure is CUHK02.Source: §27 · p. 7
- KD reduces training volatility and helps convergence, but does not guarantee higher rank-1 accuracy: improvement is reported on CUHK03-NP, while the advantage is unclear on DukeMTMC-reID.The authors suggest the public dataset’s domain distribution affects the federated model’s performance on each dataset.Source: §27 · p. 7
5.2 Weight Adjustment
- Cosine Distance Weight dynamically gives larger aggregation weights to clients whose models change more, so newly learned knowledge can better affect the federated model.Clients estimate changes using cosine distance between logits before and after local training.Source: §5.2 · p. 8
- Cosine distance weight improves performance on all datasets, and its local models outperform local-training models on every dataset.The authors interpret this as evidence that clients with different dataset sizes can obtain better-quality models through federated learning than through their individual local training.Source: §5.2 · p. 8, Table 3 · p. 8
5.3 Knowledge Distillation and Weight Adjustment
- On DukeMTMC-reID and MSMT17, FedPav with KD and cosine distance weight improves both performance and training convergence.Figure 8 evaluates these two datasets every 10 rounds; results for other datasets are reported in the supplementary material.Source: Figure 8 · p. 8, §5.3 · p. 8
6 CONCLUSION
The conclusion reports a benchmark analysis of FedReID under simulated real-world statistical heterogeneity and two optimization methods. It states that the methods improve convergence and performance, while noting that system heterogeneity remains for future work.
6 CONCLUSION
- The benchmark simulates real-world statistical heterogeneity in person re-identification and is used to analyze FedReID bottlenecks and insights.It defines federated scenarios and introduces FedPav.Source: §6 · p. 8
- KD fine-tunes the server model using knowledge from clients on a public dataset to address convergence, while aggregation weights vary with client model changes to improve large-dataset performance.The conclusion reports that the optimization methods facilitate convergence and achieve better performance.Source: §6 · p. 8
- The study focuses on statistical heterogeneity in real-world FedReID; system heterogeneity is left for future work.Source: §6 · p. 9
A.1 Experiment Settings
The appendix specifies default training settings and reports experiments varying batch size and local epochs. It states that the batch-size experiments support the paper’s conclusion, while the local-epoch comparison is described without an outcome in these passages.
A.1 Experiment Settings
- The default setup uses batch size 32, one local epoch, and 300 training rounds, with separate learning rates for the identity classifier and backbone.The learning rates are 0.05 and 0.005 respectively; server fine-tuning uses 0.0005.Source: §A.1 · p. 10
A.2 Impact of Batch Size
- Experiments compare batch sizes 32, 64, and 128 across rank-1, rank-5, rank-10, and mAP accuracy, and the authors say they demonstrate the same conclusion as the paper.Figure 9 presents the mAP comparison; Tables 4–6 report the four metrics.Source: §A.2 · p. 10
A.3 Impact of Local Epoch
- The appendix compares mAP performance for local epoch counts of 1, 5, and 10.The supplied passage identifies the comparison but does not state its outcome.Source: §A.3 · p. 10
A.4 Local Training
The supplementary results report local-training and FedPav variants, including KD and cosine distance weight (CDW). Across most datasets, the combined method has better convergence and performance, with a stated exception for peak performance on PRID2011.
A.4 Local Training
- Local-training performance is reported for rank-1, rank-5, rank-10, and mAP accuracy across the datasets.Table 7 contains the local-training results; the supplied passages give no values.Source: §A.4 · p. 10, Table 7 · p. 10
A.5 FedPav with Cosine Distance Weight
- The appendix reports rank-1, rank-5, rank-10, and mAP results for FedPav with CDW across all nine datasets.Table 8 is identified as the performance report; these passages provide no cell values.Source: §A.5 · p. 10, Table 8 · p. 11
A.6 FedPav with Knowledge Distillation
- KD facilitates convergence in FedReID training, although it does not guarantee performance improvement; rank-1 accuracy and mAP show similar patterns.Figures 11 and 12 compare FedPav with and without KD.Source: §A.6 · p. 10, Figure 11 · p. 12, Figure 12 · p. 12
A.7 FedPav with KD and CDW
- FedPav with KD and CDW achieves better convergence and superior performance on most datasets, but falls below FedPav’s fluctuating best performance on PRID2011.The authors suggest the domain gap between public dataset CUHK02 and PRID2011 could explain the slight decay, and state that careful public-dataset selection could yield better results.Source: §A.7 · p. 10, Figure 13 · p. 13, Figure 14 · p. 13
B TRAINING ALGORITHM
The appendix presents FedAvg and FedPav algorithm details, identifying FedPav as combining knowledge distillation with cosine-distance weighting.
B TRAINING ALGORITHM
- Algorithm 4 presents the Federated Averaging algorithm proposed by McMahan et al. [21].The passage says the algorithm's details are presented in this section.Source: B TRAINING ALGORITHM · p. 14
- The FedAvg algorithm comments identify n as the total dataset size and n_k as client k's dataset size.Source: B TRAINING ALGORITHM · p. 14
20 return
- Algorithm 5 presents FedPav with knowledge distillation and cosine-distance weighting.The passage identifies this as the combined algorithm, without giving further procedural details.Source: §20 · p. 14, B TRAINING ALGORITHM · p. 14