Abstract: Pedestrian re-recognition is a challenging task in the field of computer vision. This task focuses on the appearance change pattern of individuals. Due to the drastic variation of appearance feature, there is small sample problem in metric learning for person re-identification. In this paper, a transfer metric learning based method is proposed. By minimizing the difference between the distribution of source data and target data, the proposed method achieves the transform of metric model from source dataset to target dataset. The proposed method not only enhances the diversity of training samples which improves the discrimination of metric model, but also improves its generalization. Finally, the effectiveness and accuracy of the proposed method are verified on the VIPeR and CUHK01 datasets by the pre-training on iLIDS dataset.

Key words: person re-identification, metric learning, transfer learning

摘要： 行人再识别技术是计算机视觉领域中一个具有挑战性的任务。该任务针对个体的外观变化模式展开研究，特征变化剧烈，存在小样本问题，而通过提出的一种基于迁移学习的度量学习模型，可约束不同数据集样本分布的差异，实现度量模型在不同数据集上的迁移。该算法不仅增强了度量模型训练样本的多样性，提高了分辨能力，同时提升了样本的适应性。最后，通过在iLIDS数据集进行度量模型的预训练，并在VIPeR和CUHK01两个数据集上进行的迁移学习，验证了算法的有效性和准确性。

关键词: 行人再识别, 度量学习, 迁移学习