[Paper Review] Rethinking Sampling Strategies for Unsupervised Person Re-identification

Xumeng Han, Xuehui Yu+6 more|arXiv (Cornell University)|Jul 7, 2021

Video Surveillance and Tracking Methods48 references4 citations

TL;DR

This paper proposes group sampling, a novel strategy that improves unsupervised person re-identification by grouping same-class samples to enhance statistical stability and mitigate deteriorated over-fitting. It achieves state-of-the-art performance on Market-1501, DukeMTMC-reID, and MSMT17 under purely camera-agnostic settings, outperforming existing methods without additional parameters or computational cost.

ABSTRACT

Unsupervised person re-identification (re-ID) remains a challenging task. While extensive research has focused on the framework design and loss function, this paper shows that sampling strategy plays an equally important role. We analyze the reasons for the performance differences between various sampling strategies under the same framework and loss function. We suggest that deteriorated over-fitting is an important factor causing poor performance, and enhancing statistical stability can rectify this problem. Inspired by that, a simple yet effective approach is proposed, termed group sampling, which gathers samples from the same class into groups. The model is thereby trained using normalized group samples, which helps alleviate the negative impact of individual samples. Group sampling updates the pipeline of pseudo-label generation by guaranteeing that samples are more efficiently classified into the correct classes. It regulates the representation learning process, enhancing statistical stability for feature representation in a progressive fashion. Extensive experiments on Market-1501, DukeMTMC-reID and MSMT17 show that group sampling achieves performance comparable to state-of-the-art methods and outperforms the current techniques under purely camera-agnostic settings. Code has been available at https://github.com/ucas-vg/GroupSampling.

Motivation & Objective

• To investigate why sampling strategies significantly impact unsupervised person re-ID performance despite identical frameworks and loss functions.
• To identify deteriorated over-fitting as a key cause of performance degradation under random sampling.
• To propose a group sampling strategy that enhances statistical stability by grouping same-class samples for more consistent feature learning.
• To demonstrate that group sampling outperforms existing methods under purely camera-agnostic settings, where camera knowledge is not used in sampling.

Proposed method

• Group sampling collects samples from the same identity into groups, ensuring that each mini-batch contains representative clusters of the same class rather than randomly mixed samples.
• The model is trained using normalized group samples, which stabilizes feature representation by reducing the influence of individual outlier samples on class-level optimization.
• Outlier treatment is restructured to isolate unclustered outliers from grouped instances, using a dedicated outlier group (Tre. II) to prevent contamination of cluster representations.
• The method integrates seamlessly into the contrastive learning baseline by updating pseudo-label generation and representation learning through consistent class-wide trends.
• It avoids the need for triplet sampling by directly grouping same-class samples, eliminating the need for fixed K-sampling and reducing undersampling/oversampling trade-offs.
• The approach maintains intra-class similarity and inter-class distinction by preventing class-level representation collapse due to noisy or dominant individual samples.

Experimental results

Research questions

• RQ1 Why does random sampling degrade performance in unsupervised person re-ID despite identical model architecture and loss function?
• RQ2 What causes deteriorated over-fitting in unsupervised re-ID, and how can it be mitigated through sampling design?
• RQ3 How does grouping same-class samples improve statistical stability and representation quality in contrastive learning?
• RQ4 Can a simple sampling strategy like group sampling achieve performance comparable to state-of-the-art methods without additional parameters or camera knowledge?
• RQ5 Why does triplet sampling perform better than random sampling, and can a more principled group sampling strategy surpass it?

Key findings

• Group sampling achieves the best mAP on DukeMTMC-reID, outperforming all other methods by at least 1.8% under purely camera-agnostic settings.
• On Market-1501, group sampling achieves second-best performance among all methods, demonstrating strong generalization across datasets.
• The method outperforms state-of-the-art techniques on MSMT17 under purely camera-agnostic conditions, where many SOTA methods rely on camera-specific sampling.
• Replacing outlier treatment from Tre. I to Tre. II in group sampling leads to significant performance gains, confirming the importance of isolating outliers from clusters.
• Performance degrades when camera knowledge is removed from state-of-the-art methods like HCD and ICE, highlighting the superiority of group sampling in truly camera-agnostic settings.
• The ablation study confirms that group sampling stabilizes feature learning by reducing the impact of individual samples, directly countering deteriorated over-fitting.