Abstract:

When tested in another domain, the person Re-identification（ReID） model trained in a single domain has a huge performance drop, in order to solve this problem, an algorithm combining the changes of inter-domain and intra-domain is proposed. Firstly, the PR strategy is used to partition the person feature map to improve the model generalization ability. Then, for the change of the inter-domain, pose-invariance between the domains is introduced to narrow the posture gap between person in the source and target domains. Finally, for intra-domain changes, pose-invariance in a domain, sample-invariance, neighborhood-invariance and camera-invariance are introduced to increase the distance between different person and reduce the distance between the same person. Experiments demonstrate that the PR strategy and five invariance properties effectively enhance the domain adaptability in cross-domain person ReID. At the same time, the recognition accuracy is effectively improved compared with algorithms that focus only on inter-domain changes or intra-domain changes.

Key words: cross-domain person re-identification, domain adaptation, invariance

摘要：

针对单域训练的行人重识别模型迁移到另一个域内测试时性能巨大下降的问题，提出一种结合域间与域内变化的跨域行人重识别算法。采用PR策略将行人特征图进行分区处理，提高模型泛化能力。针对域间变化，引入域间姿势不变性，缩小源域和目标域行人的姿势差距。针对域内变化，引入域内姿势不变性、样本不变性、邻域不变性和相机风格不变性，扩大不同行人之间的距离，缩小相同行人之间的距离。实验表明，PR策略和5个不变性能有效地增强跨域行人重识别中的域自适应性，与只注重域间变化或域内变化的算法相比，其识别精度得到有效提升。

关键词: 跨域行人重识别, 域自适应, 不变性