Scholarlink

<p>This is an extended version of <i>DVS-StereoInsect: An Event-Based Stereo Dataset for Foreground-Background Insect Segmentation</i> presented at CAIP 2025. The conference paper presented a dataset and compared two approaches for foreground-background segmentation of event camera, or dynamic vision sensor (DVS), insect recordings. This is a necessary first step to implement event-based insect monitoring techniques, which we believe can provide valuable tools to researchers in addition to existing methods. In this work, we extend upon the results presented in the original paper by evaluating additional architectures with the goal of creating a more in-depth understanding of their influence on segmentation performance. After describing the dataset, the work presents results for a non-machine-learning-based method, three variations of <i>U-Net</i> architectures, <i>SegFormer</i>, <i>U-Net++</i>, <i>MaNet</i> and <i>LinkNet</i> models. The highest-performing model, based on the <i>U-Net</i> architecture, achieves an MCC of 0.955 detecting wasps and 0.850 detecting varied insects in front of a complex background. The evaluation of the preexisting method on the new dataset shows that it is not applicable in all use cases.</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

DVS-StereoInsect: experimental evaluation of an event-based stereo dataset for foreground-background insect segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月29日

查看全文
  • Colin Gebler,
  • Jonas Funk,
  • Regina Pohle-Fröhlich,
  • Andreas Wagner

摘要

This is an extended version of DVS-StereoInsect: An Event-Based Stereo Dataset for Foreground-Background Insect Segmentation presented at CAIP 2025. The conference paper presented a dataset and compared two approaches for foreground-background segmentation of event camera, or dynamic vision sensor (DVS), insect recordings. This is a necessary first step to implement event-based insect monitoring techniques, which we believe can provide valuable tools to researchers in addition to existing methods. In this work, we extend upon the results presented in the original paper by evaluating additional architectures with the goal of creating a more in-depth understanding of their influence on segmentation performance. After describing the dataset, the work presents results for a non-machine-learning-based method, three variations of U-Net architectures, SegFormer, U-Net++, MaNet and LinkNet models. The highest-performing model, based on the U-Net architecture, achieves an MCC of 0.955 detecting wasps and 0.850 detecting varied insects in front of a complex background. The evaluation of the preexisting method on the new dataset shows that it is not applicable in all use cases.