• 主题
高级搜索  >
历史搜索 清空历史
首页> 外文会议>International Conference on Artificial Neural Networks >Learning Camera-Invariant Representation for Person Re-identification
【24h】

Learning Camera-Invariant Representation for Person Re-identification

机译：学习相机不变表示法进行人员重新识别

获取原文
页面导航
  • 摘要
  • 著录项
  • 相似文献
  • 相关主题

摘要

Person re-identification (re-ID) problem aims to retrieve a person from an image gallery captured across multiple cameras. However, images of the same identity have variations due to the change in camera views. So learning a camera-invariant representation is one objective of re-identification. In this paper, we propose a camera-style transfer model for generating images, and a fake triplet loss for training the person feature embedding model. We train a StarGAN, a kind of generative adversarial networks, as our transfer model, which can transfer the style of an image from one camera to multiple different camera-styles by a generator network. So the image set is expanded with style-transferred images. However, style transferring yields image distortion, which misleads the training of feature embedding model. To overcome the influence of image distortion, we consider the gap between fake and real images, then we propose a fake triplet loss to capture the camera-invariant information of fake images. We do a series of experiments on the Market-1501, DukeMTMC-reID, and CUHK03 datasets, and show the effectiveness of our methods.
机译：人员重新识别（re-ID）问题旨在从多个摄像机捕获的图像库中检索人员。但是，由于摄像机视图的变化，具有相同身份的图像会有所不同。因此，学习相机不变表示是重新识别的目标之一。在本文中，我们提出了一种用于生成图像的相机式传输模型，以及一种用于训练人物特征嵌入模型的伪三元组损失。我们将一种生成型对抗网络StarGAN训练为我们的传输模型，该模型可以通过生成器网络将图像样式从一个摄像机转换为多种不同的摄像机样式。因此，图像集将使用样式转移的图像进行扩展。但是，样式转移会产生图像失真，这会误导特征嵌入模型的训练。为了克服图像失真的影响，我们考虑了假图像与真实图像之间的差距，然后提出了一个假三元损失来捕获假图像的相机不变信息。我们对Market-1501，DukeMTMC-reID和CUHK03数据集进行了一系列实验，并证明了我们方法的有效性。

著录项

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱：kefu@zhangqiaokeyan.com

京公网安备：11010802029741号 ICP备案号：京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有

  • 客服微信

  • 服务号