Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (14): 219-227.DOI: 10.3778/j.issn.1002-8331.2304-0269

• Graphics and Image Processing • Previous Articles     Next Articles

Generalized Pedestrian Re-Identification Method Based on Graph Sampling

MIN Feng, MAO Yixin, KUANG Yonggang, PENG Weiming, HAO Linlin, WU Bo   

  1. Hubei Key Laboratory of Intelligent Robot, Wuhan Institute of Technology, Wuhan 430205, China
  • Online:2024-07-15 Published:2024-07-15

图采样泛化行人重识别算法

闵锋,毛一新,况永刚,彭伟明,郝琳琳,吴波   

  1. 武汉工程大学 智能机器人湖北省重点实验室,武汉 430205

Abstract: Recent study has shown that deep feature matching methods in metric learning, combined with large-scale and diverse training data, can significantly enhance the generalization ability of person re-identification. However, many existing methods generate large memory and computational costs, such as classification parameters or class memory learning. To address these issues, a new generalization person re-identification method based on correlation graph sampler(CGS) is proposed. The basic idea of CGS is to construct a nearest neighbor relationship graph for all classes using local sensitive Hashing (LSH) and feature metrics at the beginning of training. This ensures that each small batch of training samples is composed of randomly selected base classes and near-neighboring classes that are similar to the base classes to provide informative and challenging learning examples and improve the discriminative learning ability of person re-identification models. The sampling principle of CGS is influenced by the quality of features extracted by the backbone network, and therefore, the sampling ability of CGS can be enhanced with the training of the backbone network and has learnability.Through cross-evaluation of this method on large-scale datasets(including CUHK03, Market-1501, and MSMT17), extensive experimental results demonstrate the effectiveness of this method and showcase its potential in person re-identification applications.

Key words: person re-identification, metric learning, correlation graph sampler, local sensitive Hashing (LSH)

摘要: 最近的研究表明,度量学习中的深度特征匹配方法,结合大规模、多样化的训练数据,可以显著增强人员再识别的泛化能力。然而,许多现有的方法会产生大量的内存和计算成本,如分类参数或类记忆学习等。为解决上述问题,提出了一种新的基于相关性图采样(correlation graph sampler,CGS)的泛化行人重识别算法,CGS的基本思想是在训练开始时使用局部敏感哈希函数(locality-sensitive Hashing,LSH)和特征度量为所有类构造最近邻关系图。这确保了每一小批训练样本由随机选择的基类和与基类具有相似性的近邻类组成,以提供信息量大且具有挑战性的学习示例,提高行人重识别模型的判别性学习能力。CGS的采样原理会受主干网提取的特征质量影响,因此CGS采样能力会随着主干网的训练而增强,具有可学习性。通过在大规模数据集(包括CUHK03、Market-1501和MSMT17)上交叉评估该方法,广泛的实验结果证实了该方法的有效性,并展示了其在行人重识别应用中的潜力。

关键词: 行人重识别, 度量学习, 相关性图采样, 局部敏感哈希函数