ShapeSplat
收藏资源简介:
ShapeSplat数据集由苏黎世联邦理工学院计算机视觉实验室创建,包含65,000个来自87个不同类别的3D高斯喷射对象。该数据集利用ShapeNet和ModelNet数据集构建,通过2 GPU年的计算时间生成。数据集的创建过程包括从CAD模型渲染2D图像,初始化高斯中心,并通过重要性分数进行高斯修剪。ShapeSplat数据集主要用于自监督预训练和监督微调,支持分类和3D部分分割任务,旨在解决3D视觉任务中的场景理解和编辑问题。
The ShapeSplat dataset was developed by the Computer Vision Laboratory at ETH Zurich. It contains 65,000 3D Gaussian Splatting objects belonging to 87 distinct categories. Constructed using the ShapeNet and ModelNet datasets, the dataset required 2 GPU-years of computational time to generate. Its creation process includes rendering 2D images from CAD models, initializing Gaussian centers, and performing Gaussian pruning based on importance scores. The ShapeSplat dataset is primarily used for self-supervised pre-training and supervised fine-tuning, supports classification and 3D part segmentation tasks, and aims to address scene understanding and editing challenges in 3D vision tasks.
ShapeSplat: A Large-scale Dataset of Gaussian Splats and Their Self-Supervised Pretraining
基本信息
- 标题: ShapeSplat: A Large-scale Dataset of Gaussian Splats and Their Self-Supervised Pretraining
- 作者:
- Qi Ma<sup>1,2</sup>
- Yue Li<sup>3</sup>
- Bin Ren<sup>2,4,5</sup>
- Nicu Sebe<sup>5</sup>
- Ender Konukoglu<sup>1</sup>
- Theo Gevers<sup>3</sup>
- Luc Van Gool<sup>1,2</sup>
- Danda Pani Paudel<sup>2</sup>
- 单位:
- <sup>1</sup>ETH Zürich
- <sup>2</sup>INSAIT
- <sup>3</sup>University of Amsterdam
- <sup>4</sup>University of Pisa
- <sup>5</sup>University of Trento
- 链接:
摘要
- TL;DR: 我们提出了ShapeSplat数据集和Gaussian-MAE方法,该方法能够直接在3DGS参数上进行掩码预训练。
- 主要贡献:
- ShapeSplat: 一个大规模的高斯散点数据集,包含65K个对象,涵盖87个独特类别。
- Gaussian-MAE: 基于掩码自编码器的自监督预训练方法,适用于高斯散点。
- 提出了新颖的高斯特征分组和散点池化层,这些层针对高斯参数进行了定制,能够更好地进行重建和下游任务性能提升。
演示视频
框架图
Bibtex
bibtex @misc{ma2024shapesplatlargescaledatasetgaussian, title={ShapeSplat: A Large-scale Dataset of Gaussian Splats and Their Self-Supervised Pretraining}, author={Qi Ma and Yue Li and Bin Ren and Nicu Sebe and Ender Konukoglu and Theo Gevers and Luc Van Gool and Danda Pani Paudel}, year={2024}, eprint={2408.10906}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2408.10906}, }




