Adaptive Intra-Class Variation Contrastive Learning for Unsupervised Person Re-Identification

TL;DR

This work tackles unsupervised person re-identification by introducing AdaInCV, which measures per-cluster intra-class variation to adapt memory updates. It presents two strategies, Adaptive Sample Mining (AdaSaM) and Adaptive Outlier Filter (AdaOF), within a memory-dictionary, contrastive-learning framework guided by curriculum-like adaptation and a teacher-student EMA setup. The approach yields state-of-the-art results on Market-1501 and MSMT17 without camera information and with faster convergence, highlighting the value of per-cluster adaptation and selective use of challenging samples and outliers. Overall, AdaInCV improves robustness and generalization in fine-grained Re-ID and offers a practical path to more responsive, efficient unsupervised contrastive learning systems.

Abstract

The memory dictionary-based contrastive learning method has achieved remarkable results in the field of unsupervised person Re-ID. However, The method of updating memory based on all samples does not fully utilize the hardest sample to improve the generalization ability of the model, and the method based on hardest sample mining will inevitably introduce false-positive samples that are incorrectly clustered in the early stages of the model. Clustering-based methods usually discard a significant number of outliers, leading to the loss of valuable information. In order to address the issues mentioned before, we propose an adaptive intra-class variation contrastive learning algorithm for unsupervised Re-ID, called AdaInCV. And the algorithm quantitatively evaluates the learning ability of the model for each class by considering the intra-class variations after clustering, which helps in selecting appropriate samples during the training process of the model. To be more specific, two new strategies are proposed: Adaptive Sample Mining (AdaSaM) and Adaptive Outlier Filter (AdaOF). The first one gradually creates more reliable clusters to dynamically refine the memory, while the second can identify and filter out valuable outliers as negative samples.

Figures (3)

• Figure 1: Using the t-SNE Van_2014_tSNE, the results of the initial clustering of the model are depicted in the figure on the left. 10 pseudo-labels were randomly selected from the clustering results. The features were then plotted on the graph after dimensionality reduction. The figure on the right shows the two randomly selected clusters, with the scatter points replaced by the original graphs. As depicted in the graph, variations exist in the intra-class density and the distance between the most difficult samples within clusters across different clusters. Using the same learning strategy to select samples within clusters for updating features is inaccurate. Therefore, in this work, the algorithm proposed in this paper is applicable to different clusters after clustering, rather than all clusters in the entire training set.
• Figure 2: Illustration of the framework with Adaptive Sample Mining and Adaptive Outliers Filter, which work together to provide reliable positive and negative samples for contrastive learning. We propose a training process for adaptive sample mining, which utilizes the idea of curriculum learning to enable the model to select samples of the suitable difficulty level for updating the corresponding clustering features based on the learning capacity of each cluster.
• Figure 3: Schematic diagram of Memory Bank update using the method proposed in this paper and the existing SOTA method