Scholarlink

<p>The advancement of artificial intelligence has spurred progress across diverse scientific fields, with deep learning techniques enhancing autonomous driving and vessel detection applications. The training of deep learning models relies on the construction of datasets. We present a tri-band (visible, short-wave infrared, long-wave infrared) vehicle and vessel dataset for object detection applications and multi-band image fusion. The dataset consists of thousands of images with JPG and PNG formats, and information including acquisition dates, locations, among others. The features of the dataset are time synchronization and field-of-view consistency. About 60% of the dataset has been manually labeled with object instances to train and evaluate well-established object detection algorithms. After training with YOLOv8 and SSD object detection algorithms, all models have mAP values above 0.6 at an IoU threshold of 0.5, which indicates good recognition performance for this dataset. In addition, a preliminary validation of wavelet-based multi-band image fusion was performed. As far as we know, the dataset is the first publicly available tri-band optical image 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

Tri-band vehicle and vessel dataset for artificial intelligence research

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年4月9日

查看全文
  • Yingjian Liu,
  • Gangnian Zhao,
  • Shuzhen Fan,
  • Cheng Fei,
  • Junliang Liu,
  • Zhishuo Zhang,
  • Liqian Wang,
  • Yongfu Li,
  • Xian Zhao,
  • Zhaojun Liu

摘要

The advancement of artificial intelligence has spurred progress across diverse scientific fields, with deep learning techniques enhancing autonomous driving and vessel detection applications. The training of deep learning models relies on the construction of datasets. We present a tri-band (visible, short-wave infrared, long-wave infrared) vehicle and vessel dataset for object detection applications and multi-band image fusion. The dataset consists of thousands of images with JPG and PNG formats, and information including acquisition dates, locations, among others. The features of the dataset are time synchronization and field-of-view consistency. About 60% of the dataset has been manually labeled with object instances to train and evaluate well-established object detection algorithms. After training with YOLOv8 and SSD object detection algorithms, all models have mAP values above 0.6 at an IoU threshold of 0.5, which indicates good recognition performance for this dataset. In addition, a preliminary validation of wavelet-based multi-band image fusion was performed. As far as we know, the dataset is the first publicly available tri-band optical image dataset.