基于三维点云的PIP掩码自编码器
PIP Masked Autoencoders Based on 3D Point Cloud
陈博 1袁鑫攀1
作者信息
- 1. 湖南工业大学,湖南 株洲 412007
- 折叠
摘要
现有的三维点云MAE的算法存在位置信息泄露问题和模态单一问题.为了解决这些问题,文章提出了一种用于点云-图像-点云MAE算法,称为PIP-MAE,该算法通过丰富二维图像知识来指导三维点云预训练模型,对输入的三维点云及其投影的二维图像进行随机掩模,然后重建两种模态的遮掩信息.对下游任务进行了实验,验证了PIP-MAE算法的有效性,提高了下游任务精度,能广泛用于各类下游任务.
Abstract
Existing algorithms for 3D point cloud MAE suffer from issues such as position information leakage and lack of diversity in modes.To address these problems,this paper proposes a PIP-MAE algorithm for point cloud-image-point cloud MAE.The algorithm guides the 3D point cloud pre-training model by enriching 2D image knowledge,randomly masks the input 3D point cloud and its projected 2D image,and then reconstructs the masked information for both modes.The experiments on downstream tasks validate the effectiveness of the PIP-MAE algorithm,and it improves accuracy of these downstream tasks,which can be widely used in various types of downstream tasks.
关键词
深度学习/点云重建/点云分类/点云分割Key words
Deep Learning/point cloud reconstruction/point cloud classification/point cloud segmentation引用本文复制引用
基金项目
湖南省自然科学基金(2022JJ30231)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据