Source-linked AI summary

Multiple Expert Brainstorming for Domain Adaptive Person Re-identification

Yunpeng Zhai, Qixiang Ye, Shijian Lu, Mengxi Jia, Rongrong Ji, Yonghong Tian

arXiv:2007.01546v3cs.CV

TL;DR

Domain adaptive person re-ID suffers performance drops across domains, while ensemble learning remains underexplored under unsupervised conditions. MEB-Net pre-trains heterogeneous experts on labeled source data and adapts them through mutual brainstorming with authority regularization. Experiments report superior performance over state-of-the-art and other compared methods on Market-1501 and DukeMTMC-reID.

  • Problem

    Domain adaptive person re-ID remains an open challenge because supervised models can suffer catastrophic performance drops on new domains, while model ensembles remain barely explored for this task.

  • Method

    MEB-Net pre-trains multiple differently architected expert networks on labeled source data and adapts them to an unlabeled target domain through mutual learning and authority regularization.

  • Results

    MEB-Net performs significantly better than compared single-model and baseline-ensemble methods and achieves significant gains over state-of-the-art methods on Market-1501 and DukeMTMC-reID.

  • Takeaways & Limitations

    MEB-Net assembles the discrimination capability of multiple networks while requiring only a single model during inference.

Abstract

from arXiv · show

Often the best performing deep neural models are ensembles of multiple base-level networks, nevertheless, ensemble learning with respect to domain adaptive person re-ID remains unexplored. In this paper, we propose a multiple expert brainstorming network (MEB-Net) for domain adaptive person re-ID, opening up a promising direction about model ensemble problem under unsupervised conditions. MEB-Net adopts a mutual learning strategy, where multiple networks with different architectures are pre-trained within a source domain as expert models equipped with specific features and knowledge, while the adaptation is then accomplished through brainstorming (mutual learning) among expert models. MEB-Net accommodates the heterogeneity of experts learned with different architectures and enhances discrimination capability of the adapted re-ID model, by introducing a regularization scheme about authority of experts. Extensive experiments on large-scale datasets (Market-1501 and DukeMTMC-reID) demonstrate the superior performance of MEB-Net over the state-of-the-arts.

1 Introduction

Domain adaptive person re-ID remains challenging because supervised models can suffer catastrophic performance drops in new domains. MEB-Net addresses this gap by adapting heterogeneous expert networks through mutual learning and authority regularization, with reported gains on Market-1501 and DukeMTMC-reID.

  • Domain adaptive person re-ID remains an open challenge because supervised methods can experience catastrophic performance drops on new domains.
  • Existing unsupervised approaches align feature distributions, translate images with GANs, or iteratively generate pseudo-labels through self-supervised learning and clustering.
  • MEB-Net pre-trains multiple differently architected networks as expert models, then adapts them through brainstorming-based mutual learning on unlabeled target data.
  • Authority regularization accommodates heterogeneous experts by modulating their authority according to target-domain feature distributions and enhancing re-ID discrimination.
  • MEB-Net achieves significant performance gains over state-of-the-art methods on Market-1501 and DukeMTMC-reID.

2 Related Works

Prior domain-adaptive person re-ID research emphasizes alignment, image translation, and self-supervised pseudo-labeling, while model-ensemble methods remain little explored for this task. Existing transfer and ensemble mechanisms also do not directly fit unsupervised re-ID with non-overlapping source and target identities.

  • Unsupervised domain adaptation assumes labeled source data but completely unknown target identities, motivating methods that exploit unlabeled target samples.
  • Feature distribution alignment: Feature-alignment methods reduce source–target distribution variation, including approaches based on Maximum Mean Discrepancy and shared attribute spaces.
  • Image-style transformation: GAN-based methods transfer source images toward target domains while preserving identity information and addressing camera or domain differences.
  • Self-supervised learning: Self-supervised approaches use clustering, graph matching, reciprocal search, or exemplar invariance to predict and refine target pseudo-labels.
  • Model ensemble: Model ensembles have been barely explored for domain-adaptive person re-ID despite strong performance in other tasks.
  • Knowledge transfer: Teacher–student transfer uses teacher outputs to supervise students, while mutual-learning variants exchange supervision among multiple models.
  • Knowledge transfer: Existing heterogeneous teacher–student methods largely neglect differing teacher confidence, and conventional ensemble methods cannot directly handle non-overlapping source and target label spaces.

3 The Proposed Approach

MEB-Net transfers knowledge from multiple source-pretrained expert architectures to an unlabeled target domain through iterative clustering and mutual learning. Authority regularization weights experts according to their target-domain discrimination capability.

  • Overview: MEB-Net uses labeled source data and unlabeled target data to learn a transferred person re-ID model through model ensembling.Multiple expert models are initialized from different network architectures and source-domain supervision.
  • Overview: The method pre-trains multiple heterogeneous expert networks on the source domain, then adapts them by iteratively brainstorming on target-domain samples.Source pre-training produces expert models, while target adaptation alternates clustering-based pseudo-label generation with feature learning.
  • Expert brainstorming: Mutual learning transfers knowledge through experts’ predictions and feature-based losses, while temporal averaging supplies soft labels to reduce expert homogenization.The brainstorming losses include mutual identity and mutual triplet terms; temporal averaging is used instead of current predictions for knowledge transfer.
  • Target-domain adaptation: Target-domain pseudo-labels are generated by clustering features averaged across experts, and the resulting cluster identities supervise feature learning.Each epoch averages expert features, applies mini-batch k-means, and uses cluster IDs as pseudo-labels.
  • Optimization: The overall training objective combines voting loss, mutual learning losses, and authority-weighted individual brainstorming losses across experts.Voting loss combines identity and triplet losses, and the total loss sums the individual brainstorming losses for all expert models.
  • Authority regularization: Authority regularization modulates each expert’s influence using inter-/intra-cluster scatter, giving greater authority to experts with better target-domain discrimination.The scheme is designed to accommodate heterogeneous architectures and facilitate target-domain discrimination.

4 Experiments

Experiments evaluate MEB-Net for unsupervised transfer between Market-1501 and DukeMTMC-reID, comparing it with state-of-the-art methods and testing its components. MEB-Net outperforms comparison approaches, while ablations show benefits from temporal averaging, mutual learning, authority regularization, and iterative clustering.

  • Datasets and Evaluation Metrics: MEB-Net is evaluated by using one dataset as source and the other as target, with CMC and mAP as metrics.The experiments use Market-1501 and DukeMTMC-reID for bidirectional adaptation.
  • Comparison with State-of-the-Arts: MEB-Net outperforms hand-crafted and feature-alignment approaches on the evaluated domain-adaptive person re-ID transfers.The paper attributes these gains to more discriminative deep representations and more effective use of target-domain similarities.
  • Comparison with State-of-the-Arts: MEB-Net exceeds GAN-based methods without generating new images, indicating more efficient use of unlabelled target samples.The comparison includes methods such as ECN and other GAN-based approaches.
  • Comparison with State-of-the-Arts: 76.0% mAP and 89.9% rank-1 accuracy are achieved for DukeMTMC-reID→Market1501, exceeding MMT-500 by 4.8% and 2.2%.For Market1501→DukeMTMC-reID, MEB-Net obtains 66.1% mAP and 79.6% rank-1 accuracy, exceeding the state of the art by 3.0% and 2.8%.
  • Ablation Studies: Removing temporal averaging causes drops of 5.3% mAP and 2.8% rank-1 accuracy for Market1501→DukeMTMC-reID.The paper reports that the networks tend to become homogeneous without temporally average models.
  • Ablation Studies: Removing mutual identity loss reduces mAP from 76.0% to 70.2% on Market-1501 transfer and from 66.1% to 60.4% on DukeMTMC-reID transfer.The mutual triplet-loss ablation also lowers mAP to 74.9% and 63.0% for the two transfer directions.
  • Discussion: MEB-Net performs significantly better than single-model transfer and baseline ensemble methods, supporting mutual learning as an effective ensemble strategy.The baseline ensemble improves over single-model transfer through more accurate pseudo-labels, while MEB-Net adds mutual learning during adaptation.
  • Discussion: Performance improves during early iterative clustering and converges after 20 epochs on both datasets.The epoch evaluation reports that all networks ascend to a stable value after 20 epochs.

5 Conclusion

MEB-Net uses multiple heterogeneous expert networks, mutual learning, and authority regularization for domain-adaptive person re-ID. Experiments show improved discrimination while requiring only one expert during inference.

  • MEB-Net pre-trains multiple architectural experts and adapts them through brainstorming-based mutual learning in the target domain.Authority regularization addresses expert heterogeneity during adaptation.
  • MEB-Net improves the discrimination ability of re-ID models by assembling capabilities from multiple networks.Only a single model is required during inference.
Loading 2007.01546v3…