高级检索
    戴臣超, 王洪元, 倪彤光, 陈首兵. 基于深度卷积生成对抗网络和拓展近邻重排序的行人重识别[J]. 计算机研究与发展, 2019, 56(8): 1632-1641. DOI: 10.7544/issn1000-1239.2019.20190195
    引用本文: 戴臣超, 王洪元, 倪彤光, 陈首兵. 基于深度卷积生成对抗网络和拓展近邻重排序的行人重识别[J]. 计算机研究与发展, 2019, 56(8): 1632-1641. DOI: 10.7544/issn1000-1239.2019.20190195
    Dai Chenchao, Wang Hongyuan, Ni Tongguang, Chen Shoubing. Person Re-Identification Based on Deep Convolutional Generative Adversarial Network and Expanded Neighbor Reranking[J]. Journal of Computer Research and Development, 2019, 56(8): 1632-1641. DOI: 10.7544/issn1000-1239.2019.20190195
    Citation: Dai Chenchao, Wang Hongyuan, Ni Tongguang, Chen Shoubing. Person Re-Identification Based on Deep Convolutional Generative Adversarial Network and Expanded Neighbor Reranking[J]. Journal of Computer Research and Development, 2019, 56(8): 1632-1641. DOI: 10.7544/issn1000-1239.2019.20190195

    基于深度卷积生成对抗网络和拓展近邻重排序的行人重识别

    Person Re-Identification Based on Deep Convolutional Generative Adversarial Network and Expanded Neighbor Reranking

    • 摘要: 行人重识别任务旨在识别不相交摄像头视图下的相同行人.这项任务极具挑战性,尤其是当数据集中每个行人仅仅有几张图片时.针对行人重识别数据集中行人图片数量不足的问题,提出一个从原始数据集中生成额外训练数据的方法.在这项工作之中存在2个挑战:1)如何从原始数据集之中获取更多的训练数据;2)如何处理这些新生成的训练数据.使用深度卷积生成对抗网络来生成额外的无标签行人图片,并采用标签平滑正则化来处理这些新生成的无标签行人图片.为了进一步提升行人重识别准确度,提出了一种新的无监督重排序框架.此框架既不需要为每组图像对重新计算新的排序列表,也不需要任何人工交互或标签信息.在Market-1501,CUHK03和DukeMTMC-reID数据集上的实验验证了所提方法的有效性.

       

      Abstract: Person Re-Identification (Re-ID) focuses on identifying the same person among disjoint camera views. This task is highly challenging, especially when there exists only several images per person in the database. Aiming at the problem of insufficient number of person images in person re-identification dataset, a method that generates extra training data from the original dataset is proposed. There are two challenges in this work, one is how to get more training data from the original training set, and the other is how to deal with these newly generated training data. The deep convolutional generative adversarial network is used to generate extra unlabeled person images and label smoothing regularization is used to process these newly generated unlabeled person images. In order to further improve the accuracy of person re-identification, a new unsupervised reranking framework is proposed. This framework neither requires to recalculate a new sorted list for each image pairs nor requires any human interaction or label information. Experiments on the datasets Market-1501, CUHK03, and DukeMTMC-reID verify the effectiveness of the proposed method.

       

    /

    返回文章
    返回