Feat2GS
收藏资源简介:
Feat2GS数据集是由西湖大学和马克斯·普朗克研究所等机构创建的,旨在评估视觉基础模型(VFMs)在3D几何和纹理感知方面的能力。该数据集通过从无姿态图像中提取的VFM特征,使用轻量级的读出层将像素级特征转换为3D高斯分布,从而进行新颖视图合成任务。数据集的创建过程涉及从多个视角的图像中提取特征,并通过多视角光度损失进行训练。Feat2GS数据集主要应用于3D视觉任务,如新颖视图合成,旨在解决VFM在3D感知方面的局限性问题。
The Feat2GS dataset was developed by institutions including Westlake University and the Max Planck Institute, with the objective of evaluating the capabilities of visual foundation models (VFMs) in 3D geometry and texture perception. This dataset leverages VFM features extracted from unposed images, and employs a lightweight readout layer to convert pixel-level features into 3D Gaussian distributions for novel view synthesis tasks. The dataset creation process involves extracting features from multi-view images and training with multi-view photometric loss. The Feat2GS dataset is primarily applied to 3D vision tasks such as novel view synthesis, and is designed to address the limitations of VFMs in 3D perception.




