[Paper Review] Neighbour Consistency Guided Pseudo-Label Refinement for Unsupervised Person Re-Identification

De Cheng, Haichun Tai+3 more|arXiv (Cornell University)|Nov 30, 2022

Video Surveillance and Tracking Methods4 citations

TL;DR

This paper proposes Neighbour Consistency Guided Pseudo-Label Refinement (NCPLR), a simple yet effective method for unsupervised person re-identification that refines noisy pseudo-labels by aggregating predictions from nearest neighbors using feature-space similarity weights. The approach improves model robustness and achieves state-of-the-art performance, boosting mAP by up to 1.48% across five datasets compared to baseline clustering-based methods.

ABSTRACT

Unsupervised person re-identification (ReID) aims at learning discriminative identity features for person retrieval without any annotations. Recent advances accomplish this task by leveraging clustering-based pseudo labels, but these pseudo labels are inevitably noisy which deteriorate model performance. In this paper, we propose a Neighbour Consistency guided Pseudo Label Refinement (NCPLR) framework, which can be regarded as a transductive form of label propagation under the assumption that the prediction of each example should be similar to its nearest neighbours'. Specifically, the refined label for each training instance can be obtained by the original clustering result and a weighted ensemble of its neighbours' predictions, with weights determined according to their similarities in the feature space. In addition, we consider the clustering-based unsupervised person ReID as a label-noise learning problem. Then, we proposed an explicit neighbour consistency regularization to reduce model susceptibility to over-fitting while improving the training stability. The NCPLR method is simple yet effective, and can be seamlessly integrated into existing clustering-based unsupervised algorithms. Extensive experimental results on five ReID datasets demonstrate the effectiveness of the proposed method, and showing superior performance to state-of-the-art methods by a large margin.

Motivation & Objective

• To address the issue of noisy pseudo-labels in clustering-based unsupervised person re-identification (ReID), which degrade model performance due to label memorization and overfitting.
• To improve training stability and robustness by modeling unsupervised ReID as a label-noise learning problem.
• To develop a lightweight, plug-in method that enhances existing clustering-based USL ReID frameworks without requiring auxiliary networks or complex architectures.
• To leverage local neighborhood structure in feature space to refine pseudo-labels through transductive label propagation principles.

Proposed method

• The refined pseudo-label for each training instance is computed as a weighted ensemble of its k-nearest neighbors’ predictions, where weights are based on cosine similarity in the feature space.
• The method combines the original clustering-based pseudo-label and the neighbor-ensemble prediction via a learnable interpolation coefficient α.
• A novel neighbor consistency regularization (NCR) is introduced to penalize prediction divergence from the weighted average of neighbors’ predictions, improving training stability.
• The NCR loss is computed using a similarity graph built from feature embeddings, encouraging local consistency without relying on predicted labels.
• The framework is compatible with standard metric learning objectives like triplet loss and InfoNCE loss, and can be integrated into existing USL ReID pipelines with minimal modification.
• Hyperparameters α, ρ (neighborhood radius), and λ1, λ2 are tuned to balance refinement and supervision, with optimal values found at α=0.2, ρ=0.2, λ1=1.0, λ2=1.0.

Experimental results

Research questions

• RQ1 Can neighbor-based label refinement improve the quality of pseudo-labels in unsupervised person ReID without requiring auxiliary models or complex architectures?
• RQ2 How does neighbor consistency regularization reduce model overfitting to noisy pseudo-labels during training?
• RQ3 To what extent does the proposed NCPLR method improve mAP and Rank-1 accuracy compared to baseline clustering-based USL ReID methods?
• RQ4 What is the optimal configuration of hyperparameters α, ρ, λ1, and λ2 for maximizing performance across diverse ReID datasets?
• RQ5 Does the method remain effective when clustering quality is low, or does it depend on a relatively clean initial pseudo-label set?

Key findings

• NCPLR improves mAP by an average of 1.48% over the baseline method Lcc + Lce across five standard ReID datasets.
• When combined with a weighted refinement strategy (Eq. 7), an additional 0.48% mAP gain is achieved on average.
• The method achieves a 0.64% mAP improvement on average when applied to teacher-student networks, demonstrating its effectiveness in self-distillation settings.
• The optimal hyperparameter configuration is α=0.2, ρ=0.2, λ1=1.0, λ2=1.0, as determined through ablation on Market1501.
• The model performs best when α=0.2, indicating that a balance between original clustering and neighbor consensus yields superior results.
• The method is robust and generalizable, showing consistent gains across multiple datasets including Market1501, DukeMTMC, CUHK03, MSMT1233, and CUHK-SYSU.