Weak-edge sample extension for enhancing unsupervised feature learning

Unsupervised person re-identification (Re-ID) methods aim to learn discriminative features from unlabeled datasets for person retrieval. Most existing methods rely on clustering to generate pseudo-labels. However,due to inaccurate feature representations and inherent limitations of clustering algorithms,clustering may sometimes merge different identities into the same group. Recent methods mitigate the issue by improving feature representations or refining pseudo-labels,but they overlook critical samples located on the sparse edges of clusters,which we refer to as weak-edge samples. In this paper,we propose a Weak-edge Sample Extension (WSE) framework to learn a more accurate decision boundary by performing sample extension specifically on weak-edge samples. To achieve this,we introduce an edge strength scoring mechanism based on the neighborhood relations of individual samples within their clusters. The neighborhood structure of each sample is represented by its k-nearest neighbor clusters,and a global neighborhood frequency distribution is constructed by aggregating these local neighborhoods across the entire cluster. Based on this distribution,we compute an edge strength score to automatically detect weak-edge samples and apply sample extension,providing structural support that enhances discriminative feature learning and improves model performance. Extensive experimental results on Market-1501 and MSMT17 demonstrate the effectiveness of the proposed method.