Scholarlink

Abstract <p>To improve the person re-identification accuracy in videos, a method is proposed for finding the most effective values of hyperparameters of convolutional neural networks when calculating features of images of people. The technique consists of two stages. At the first stage, the size of the image batch and the learning rate are determined at which the maximum accuracy is achieved for the basic re-identification algorithm. At the second stage, various options for adjusting hyperparameters during the training process are investigated, making it possible to increase the accuracy relative to the maximum accuracy values obtained at the first stage. To improve the accuracy of re-identification of people, data augmentation, an extended training sample based on existing sets, and an analysis of local features of images of people are used, which allows replacing descriptors of hidden areas of the human figure. The application of the proposed approach makes it possible to increase the accuracy of person re-identification for CNN ResNet-50 and DenseNet-121 by 27.33 according to the Rank1 metric, by 31.01 according to the mAP metric, and by 33.97 according to the mINP metric relative to the basic re-identification algorithm. A comparison of the proposed algorithm with existing ones is presented when testing on the Occluded-Duke dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Two-Stage Hyperparameter Tuning and Local Feature Extraction for Improved Person Re-Identification

  • 期刊论文

  • 发表日期：

    2025年9月1日

查看全文
  • Huafeng Chen,
  • S. Ihnatsyeva,
  • R. Bohush,
  • Shiping Ye,
  • S. Ablameyko

摘要

Abstract

To improve the person re-identification accuracy in videos, a method is proposed for finding the most effective values of hyperparameters of convolutional neural networks when calculating features of images of people. The technique consists of two stages. At the first stage, the size of the image batch and the learning rate are determined at which the maximum accuracy is achieved for the basic re-identification algorithm. At the second stage, various options for adjusting hyperparameters during the training process are investigated, making it possible to increase the accuracy relative to the maximum accuracy values obtained at the first stage. To improve the accuracy of re-identification of people, data augmentation, an extended training sample based on existing sets, and an analysis of local features of images of people are used, which allows replacing descriptors of hidden areas of the human figure. The application of the proposed approach makes it possible to increase the accuracy of person re-identification for CNN ResNet-50 and DenseNet-121 by 27.33 according to the Rank1 metric, by 31.01 according to the mAP metric, and by 33.97 according to the mINP metric relative to the basic re-identification algorithm. A comparison of the proposed algorithm with existing ones is presented when testing on the Occluded-Duke dataset.