Part-based Pseudo Label Refinement for Unsupervised Person Re-identification (2203.14675v1)

Abstract: 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.

• The paper introduces a novel cross agreement score to refine global pseudo-labels by integrating both global and part-level features.
• The approach achieves a 3.8% mAP increase on Market-1501, effectively reducing label noise in unsupervised re-identification.
• The paper offers a computationally efficient framework that bypasses auxiliary networks and encourages further research on hybrid feature utilization.

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

The paper "Part-based Pseudo Label Refinement for Unsupervised Person Re-identification" presents an innovative approach to address the challenge of learning discriminative representations for person retrieval from unlabeled data. The authors propose a Part-based Pseudo Label Refinement (PPLR) framework that leverages both global and part-level features to refine pseudo-labels, thereby reducing label noise inherent in unsupervised person re-identification tasks.

Unsupervised person re-identification (re-ID) typically involves the generation of pseudo-labels from unlabeled data, which are utilized to guide the training of retrieval models. However, these pseudo-labels often suffer from noise that deteriorates performance. Existing pseudo-label refinement methods predominantly consider only global features while neglecting fine-grained details, which are crucial for this task. PPLR effectively incorporates fine-grained local context by leveraging part features, providing a robust solution to the noise issue without the computational overhead of auxiliary networks commonly used in peer-teaching approaches.

Methodology and Results

The PPLR framework introduces a novel cross agreement score that assesses the similarity of $k$-nearest neighbors between global and part features. By evaluating this complementary relationship, PPLR refines global pseudo-labels by aggregating predictions from part features. Additionally, the cross agreement score facilitates agreement-aware label smoothing (AALS) to adjust label distributions according to the alignment of part features with global pseudo-labels. The framework is elegant in its simplicity, employing self-ensemble strategies that avoid the complexities and resource demands of multiple-model approaches.

The empirical evaluation, conducted on prominent person re-ID datasets such as Market-1501 and MSMT17, demonstrates significant improvements over existing state-of-the-art methods. For instance, PPLR achieves a notable 3.8% increase in mean Average Precision (mAP) on Market-1501 over prior leading techniques. This performance gain illustrates the efficacy of integrating local context through part features into the refinement of pseudo-labels, validating the framework's design choices.

Implications and Future Research

The development and application of PPLR bear significant implications for the field of unsupervised learning in re-ID tasks. Practically, the framework advances the state-of-the-art by offering a more computationally efficient method that circumvents the need for auxiliary models while successfully addressing the persistent challenge of label noise. Theoretically, PPLR underscores the value of examining and exploiting the interplay between part-level and global features in refining pseudo-labels, prompting further investigation into hybrid feature utilization.

Future research could build upon this framework by exploring the integration of semantic recognition or parsing to enhance the semantic coherence across part features, thereby addressing potential misalignment issues observed with uniformly partitioned regions. Additionally, the principles underlying PPLR could be extended across other unsupervised learning domains, where label noise is a known hindrance.

This research opens new avenues for refining pseudo-labels in data-intensive tasks, contributing valuable insights into the integration of granular feature information into unsupervised learning frameworks. The proposed methodology not only advances person re-identification but also sets a precedent for refining learning processes with minimalized computational overhead.