Scholarlink

<p>Optical flow is a vector field that describes the displacement relationship of pixels between two images. In the real world, obtaining ground truth data for optical flow is an extremely challenging task. For a long time, researchers mainly relied on data generated from simulated environments to train optical flow models. However, this method often makes lightweight small optical flow models difficult to achieve sufficient optical flow accuracy. By proposing a knowledge distillation-based method, the accuracy of lightweight optical flow models in the real world can be improved. Additionally, a multi-teacher dataset creation method based on cycle consistency is proposed, which generates highly credible pseudo-labels for training or optimization.</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

Design of Optical Flow Model Optimization Methods in the Absence of Ground Truth

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Jijia Fang,
  • Danping Zou,
  • Changchun Pan

摘要

Optical flow is a vector field that describes the displacement relationship of pixels between two images. In the real world, obtaining ground truth data for optical flow is an extremely challenging task. For a long time, researchers mainly relied on data generated from simulated environments to train optical flow models. However, this method often makes lightweight small optical flow models difficult to achieve sufficient optical flow accuracy. By proposing a knowledge distillation-based method, the accuracy of lightweight optical flow models in the real world can be improved. Additionally, a multi-teacher dataset creation method based on cycle consistency is proposed, which generates highly credible pseudo-labels for training or optimization.