PaperlayerView exact revision on arXiv

Source-linked AI summary

Part-based Pseudo Label Refinement for Unsupervised Person Re-identification

Yoonki Cho, Woo Jae Kim, Seunghoon Hong, Sung-Eui Yoon

arXiv:2203.14675v1cs.CV

TL;DR

Unsupervised person re-ID is challenged by noisy pseudo-labels and insufficient use of fine-grained local context. PPLR refines global and part pseudo-labels using reliable cross-feature agreement, and experiments show effectiveness against prior state-of-the-art methods on several benchmarks.

  • Problem

    Noisy pseudo-labels and neglected fine-grained local context limit unsupervised person re-ID.

  • Method

    PPLR uses a cross agreement score between global and part feature neighborhoods to guide part-based pseudo-label refinement.

  • Results

    PPLR outperforms prior state-of-the-art methods on several benchmarks, with refinement gains demonstrated on Market-1501 and MSMT17.

  • Takeaways & Limitations

    Reliable complementary information from global and part features helps reduce label noise while incorporating local context into unsupervised re-ID.

  • Takeaways & Limitations

    Uniform feature-map partitioning can misalign corresponding part indices because not all images are aligned to a human body.

Abstract

from arXiv · show

Unsupervised person re-identification (re-ID) aims at learning discriminative representations for person retrieval from unlabeled data. Recent techniques accomplish this task by using pseudo-labels, but these labels are inherently noisy and deteriorate the accuracy. To overcome this problem, several pseudo-label refinement methods have been proposed, but they neglect the fine-grained local context essential for person re-ID. In this paper, we propose a novel Part-based Pseudo Label Refinement (PPLR) framework that reduces the label noise by employing the complementary relationship between global and part features. Specifically, we design a cross agreement score as the similarity of k-nearest neighbors between feature spaces to exploit the reliable complementary relationship. Based on the cross agreement, we refine pseudo-labels of global features by ensembling the predictions of part features, which collectively alleviate the noise in global feature clustering. We further refine pseudo-labels of part features by applying label smoothing according to the suitability of given labels for each part. Thanks to the reliable complementary information provided by the cross agreement score, our PPLR effectively reduces the influence of noisy labels and learns discriminative representations with rich local contexts. Extensive experimental results on Market-1501 and MSMT17 demonstrate the effectiveness of the proposed method over the state-of-the-art performance. The code is available at https://github.com/yoonkicho/PPLR.

1. Introduction

Unsupervised person re-ID avoids costly annotations but relies on noisy pseudo-labels that limit performance. PPLR uses reliable complementarity between global and part features to refine labels without auxiliary networks.

  • Unsupervised person re-ID learns discriminative retrieval features from unlabeled data, addressing the annotation cost of supervised approaches.
  • Pseudo-label methods alternate label generation and model training, but noisy labels significantly hinder unsupervised re-ID performance.
  • Existing peer-teaching refinement reduces label noise but requires multiple teacher backbones, increasing computational costs.
  • PPLR uses part features in a self-teaching framework because local texture can provide robust identity cues under pose and viewpoint variation.
  • PPLR identifies reliable global–part complementarity with a cross agreement score based on similarity between their k-nearest neighbors.

2. Related Work

Related work spans noisy-label learning, part-based re-ID, and unsupervised re-ID. PPLR differs by using part features to reduce noise in global feature clustering.

  • Learning with noisy labels: Noise-robust learning methods adjust losses or reweight samples, but several such losses target simple image classification rather than open-set person re-ID.
  • Part-based approaches for person re-ID: Part-based re-ID methods exploit fine-grained body information through parsing, attention, pose estimation, or multiple granularities.
  • Part-based approaches for person re-ID: Unlike prior part-feature methods, PPLR uses part features to reduce label noise in global feature clustering through fine-grained information.
  • Unsupervised approaches for person re-ID: Unsupervised re-ID research includes domain adaptation using labeled source data and unsupervised learning without an external labeled source domain.

3. Method

PPLR alternates global-feature clustering with training that refines pseudo-labels using reliable complementary information from global and part features. It uses cross agreement to calibrate part labels and aggregate reliable part predictions into global labels, reducing the influence of noisy pseudo-labels without auxiliary networks.

  • Overall framework: PPLR alternates clustering and training stages, clustering global features before refining pseudo-labels during model training.The clustering stage assigns global pseudo-labels and computes cross agreement; the training stage uses refined labels.
  • Part-based representation: The framework extracts a shared representation, global features by global average pooling, and part features by pooling separate feature regions.Global and part features jointly represent each image, unlike approaches using only global features.
  • Pseudo-label generation: DBSCAN clustering on global features provides one-hot pseudo-labels shared by global and part feature spaces.These pseudo-labels supervise cross-entropy and softmax-triplet training for the respective feature classifiers.
  • Training objective: The baseline objective combines global cross-entropy, part cross-entropy, softmax-triplet, and optionally inter-camera contrastive losses.The camera-aware loss pulls together same-cluster proxies from different cameras to reduce intra-class variance caused by disjoint views.
  • Cross Agreement: The cross agreement score is the Jaccard similarity between global and part k-nearest-neighbor lists, identifying whether complementary information is reliable.This addresses unreliable part information from occlusions, backgrounds, or differing semantic content.
  • Pseudo Label Refinement: Agreement-aware label smoothing trains low-agreement part features toward a uniform distribution, while part-guided refinement weights part predictions by agreement when refining global labels.The global refinement is self-teaching and avoids the additional teacher networks used by some prior methods.

4. Experiments

Experiments on Market-1501, MSMT17, and VeRi-776 evaluate PPLR through component ablations, parameter analysis, qualitative visualizations, and comparisons with prior methods. The results support both refinement components and show gains over state-of-the-art systems.

  • Ablation Study: On Market-1501, combining AALS and PGLR improves baseline mAP by 6.5% with Lcam and 8.0% without Lcam.Each refinement component improves performance, with the combined model producing the largest gains.
  • Ablation Study: AALS significantly outperforms vanilla label smoothing and part-to-part refinement by weighting smoothing according to part-label reliability.Part-to-part refinement uses less reliable complementary information because parts have less-overlapped receptive fields, whereas vanilla smoothing ignores part characteristics.
  • Ablation Study: PGLR significantly outperforms alternative refinement methods, showing that cross-agreement-weighted part predictions produce more effective global pseudo-labels.The refined labels capture fine-grained information unavailable from global features alone.
  • Parameter Analysis: Large k values create more false top-k matches and lower cross-agreement scores, while the selected parameters are k = 20 and β = 0.5.False matches make complementary-reliability estimation less effective and limit performance.
  • Qualitative Analysis: Cross-agreement visualizations assign low scores to occluded or misaligned parts and high scores to discriminative parts.When all parts are occluded, predictions collapse toward a uniform distribution and provide meaningless training signals.
  • Qualitative Analysis: AALS produces better-distributed part embeddings, while PGLR makes global-feature attention cover more diverse discriminative regions.Without AALS, embeddings overfit to identity labels; without PGLR, Grad-CAM focuses mainly on the most discriminative parts.
  • Comparison with State-of-the-Arts: PPLR exceeds RLCC by 3.8% mAP on Market-1501, 3.5% on MSMT17, and 2.0% on VeRi-776.Against ICE, PPLR gains 2.1% mAP on Market-1501 and 3.3% on MSMT17.

5. Discussion

PPLR has practical limitations related to uniformly partitioned part features and the ethical risks of re-identification technology. Misalignment can reduce the reliability of part-based information, while surveillance misuse may threaten privacy.

  • Part-based limitation: Uniform feature-map partitioning may assign the same part index to different human body regions across unaligned images.The paper identifies semantic matching or human parsing as a promising alternative.
  • Part-based limitation: Misaligned part features can produce noisy ranked lists and low cross agreement scores.This limits the reliability of complementary information between global and part features.
  • Ethical limitation: Re-identification technology may infringe privacy when surveillance systems are abused.The paper urges researchers and users to consider ethical issues and notes that DukeMTMC-reID was taken down for such concerns.

6. Conclusion

The paper concludes that PPLR uses global and local context to reduce pseudo-label noise through reliable complementary information. Experiments and ablations support its effectiveness and show performance above prior state-of-the-art methods on several benchmarks.

  • Conclusion: PPLR refines pseudo-labels in global and part feature spaces using their complementary relationship.The framework introduces a cross agreement score, agreement-aware label smoothing, and part-guided label refinement.
  • Conclusion: PPLR uses part features in a self-ensemble framework without auxiliary networks.This approach addresses label noise while incorporating local context into representation learning.
  • Conclusion: Extensive ablation studies and experiments on several benchmarks validate PPLR's effectiveness and performance over prior state-of-the-art methods.The evaluation includes qualitative analysis and ablation studies.

A.1. Parameter Analysis

The number of parts affects baseline performance because more parts provide smaller receptive fields and less contextual information. PPLR continues to improve the baseline as the number of parts changes.

  • Number of parts: Larger numbers of parts reduce each part feature's receptive field and limit the clues available for person re-identification.The resulting part features are trained with hard pseudo-labels that do not consider each part's context.
  • Number of parts: Baseline performance decreases as the number of parts increases.The paper attributes this trend to smaller receptive fields and hard pseudo-label training.
  • Number of parts: PPLR consistently improves the baseline across the analyzed numbers of parts.

A.2. Training Computational Cost

PPLR is more computationally efficient than auxiliary-teacher approaches because it uses features from a single backbone. This efficiency remains after including cross agreement score computation.

  • Computational cost: PPLR uses a single backbone and is more efficient than MEB-Net2 and MMT.The comparison considers training parameters, training-stage time, and clustering-stage time.
  • Computational cost: PPLR remains more efficient even when the time required to compute the cross agreement score is included.

A.3. Qualitative Results

The qualitative analysis links cross agreement to the reliability of part information and describes camera-aware proxy training as a way to reduce cross-camera intra-class variation.

  • Qualitative Results: Low cross agreement often indicates occlusions, backgrounds, or multiple people that make corresponding part information less discriminative.High-scoring images are described as better aligned with discriminative information, while some misaligned discriminative parts remain failure cases.
  • Qualitative Results: Figure 7 analyzes the parameter Np on Market-1501 against a part-based unsupervised re-ID baseline.The caption identifies Baseline as the part-based unsupervised re-ID framework in Section 3.1.
  • Qualitative Results: Table 4 compares training costs on Market-1501 using clustering-stage time and sec/iter for the training stage.Clustering-stage time includes feature extraction, clustering, and cross agreement score computation; sec/iter accounts for differing iterations per epoch.
  • Camera-aware Proxy Details: PPLR optionally uses camera-aware proxies when camera labels are available, computing each proxy as the centroid of features sharing a camera label and cluster.The proxy is formed from the feature subset indexed by matching camera and pseudo-label values.
  • Camera-aware Proxy Details: Inter-camera contrastive loss uses same-label proxies from different cameras as positives and nearest different-label proxies as hard negatives.Applying this loss across feature spaces is intended to reduce large intra-class variance caused by disjoint camera views.
  • Qualitative Results: Figure 8 visualizes low-50 and top-50 cross agreement cases on Market-1501 after excluding very similar and duplicated images.The visualization is intended to show varied cases rather than duplicated examples.

C. More Implementation Details

The implementation uses PyTorch with specified GPU resources and clustering settings, while Figures 9 and 10 provide enlarged versions of earlier figures.

  • More Implementation Details: The framework is implemented in PyTorch, trained on four NVIDIA TITAN RTX GPUs, and tested using one GPU.These hardware settings are stated for the implementation and evaluation procedure.
  • More Implementation Details: Clustering uses Jaccard distance with k-reciprocal encoding and k=30.The clustering distance is based on the cited k-reciprocal encoding procedure.
  • More Implementation Details: DBSCAN uses a minimum core-point neighbor count of 4 and distance thresholds of 0.7 for MSMT17 and VeRi-776 and 0.6 for Market-1501.The threshold differs by dataset, while the minimum-neighbor setting remains fixed.
  • More Implementation Details: Figures 9 and 10 are presented as enlarged versions of Figures 2 and 5, respectively.The captions identify these figures as large-format versions rather than new analyses.
Loading 2203.14675v1…