Adaptive Scheme of Clustering-Based Unsupervised Learning for Person Re-identification

Clustering and cluster-level contrastive learning are widely used functions in unsupervised re-identification methods, with a focus on extracting robust and distinctive features from data without annotations. However, existing approaches often overlook the correlation of the re-ID modules related hyperparameters with the considerable shrinkage in cluster density. This issue potentially lead to misaligned cluster representation vectors taken into account of computing cluster-level contrastive loss. Therefore, it might hinder the model's performance. To address this problem, we propose a novel method called Adaptive Scheme of Clustering-based Unsupervised Learning (ASCUL). In contrast to approaches that rely on predefined clustering hyperparameters, we incorporate a regulator executing adaptive adjustments to maximize the number of informative samples during training. Furthermore, our scheme for mining cluster representations adapts dynamically to substantial changes in intra-class variations, efficiently utilizing the clusterwise loss. Experiments on two benchmark datasets consistently show that our new approach outperforms state-of-the-art unsupervised person re-ID methods.