Scholarlink

<p>Image-to-point cloud (I2P) registration aims to align 2D images and 3D point clouds within the same scene. However, existing methods neglect the differences in learning difficulties among different samples, failing to bridge the image-to-point cloud modality gap adequately. Moreover, the features of images or point clouds within the same modality are easily affected by various geometric transformations, which further increases the complexity of the registration task. To address these issues, we propose a similarity curriculum learning framework for I2P registration, termed <b>CurrI2P</b>, which not only effectively captures similarities between different data pairs but also gradually adapts the model to more complex data transformations. Specifically, we first present inter-modality similarity curriculum learning to progressively enhance cross-modality similarity to reduce inaccurate matches caused by coarse registration at the early training stage. In addition, we design intra-modality similarity curriculum learning, particularly targeting to learn invariant features of images or point clouds within the same modality under the geometric disturbance of increasing magnitude. In experiments, we incorporate CurrI2P into different baselines to validate the feasibility and achieve state-of-the-art performance on both KITTI and nuScenes datasets. The code is available at <a href="https://github.com/lin-liwei/CurrI2P">https://github.com/lin-liwei/CurrI2P</a>.</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

CurrI2P: inter- and intra-modality similarity curriculum learning for image-to-point cloud registration

  • 期刊论文

  • 发表日期：

    2025年4月7日

查看全文
  • Liwei Lin,
  • Chunyu Lin,
  • Lang Nie,
  • Shujuan Huang,
  • Yao Zhao

摘要

Image-to-point cloud (I2P) registration aims to align 2D images and 3D point clouds within the same scene. However, existing methods neglect the differences in learning difficulties among different samples, failing to bridge the image-to-point cloud modality gap adequately. Moreover, the features of images or point clouds within the same modality are easily affected by various geometric transformations, which further increases the complexity of the registration task. To address these issues, we propose a similarity curriculum learning framework for I2P registration, termed CurrI2P, which not only effectively captures similarities between different data pairs but also gradually adapts the model to more complex data transformations. Specifically, we first present inter-modality similarity curriculum learning to progressively enhance cross-modality similarity to reduce inaccurate matches caused by coarse registration at the early training stage. In addition, we design intra-modality similarity curriculum learning, particularly targeting to learn invariant features of images or point clouds within the same modality under the geometric disturbance of increasing magnitude. In experiments, we incorporate CurrI2P into different baselines to validate the feasibility and achieve state-of-the-art performance on both KITTI and nuScenes datasets. The code is available at https://github.com/lin-liwei/CurrI2P.