Streaming3D
收藏资源简介：
Streaming3D数据集是为Streaming3D基准测试提供的资产集合。当前发布的版本为GSO30子集，这是一个包含30个常见家居物品（如闹钟、背包、积木、大象、鞋子、沙发、茶壶等）的3D对象数据集，其对象源自Google Scanned Objects。每个对象均提供完整的3D资产：包括原始网格文件（.glb, .obj）、材质文件（.mtl）和纹理（.png）。数据集的核心是为3D重建与视图合成任务构建的标准化训练与评估流水线。具体包含两个主要渲染集：1) `render_spiral_100`：提供围绕对象的100个螺旋视角的RGB图像（000.png-099.png）、对应的对象掩码（masks/）以及详细的相机变换参数（transforms.json和.npy文件），用于模型训练或重建。2) `render_mvs_25`：提供经过归一化处理的网格模型（model_norm.glb）以及25个固定视角的相机参数（.npy文件）和对应的参考渲染图（.png文件），用于模型性能的定量评估。数据集的使用协议明确：使用`render_spiral_100`中的所有图像进行训练或重建，然后使用`render_mvs_25`中的归一化模型和25个相机位姿进行评估和对比。
The Streaming3D dataset is an asset collection provided for the Streaming3D benchmark. The current released version is the GSO30 subset, which is a 3D object dataset containing 30 common household items (such as alarm clocks, backpacks, building blocks, elephants, shoes, sofas, teapots, etc.), with objects sourced from Google Scanned Objects. Each object provides complete 3D assets: including original mesh files (.glb, .obj), material files (.mtl), and textures (.png). The core of the dataset is to build a standardized training and evaluation pipeline for 3D reconstruction and view synthesis tasks. Specifically, it includes two main rendering sets: 1) `render_spiral_100`: provides 100 spiral-view RGB images around the object (000.png-099.png), corresponding object masks (masks/), and detailed camera transformation parameters (transforms.json and .npy files), for model training or reconstruction. 2) `render_mvs_25`: provides a normalized mesh model (model_norm.glb), along with 25 fixed-view camera parameters (.npy files) and corresponding reference renderings (.png files), for quantitative evaluation of model performance. The dataset usage protocol is clear: use all images in `render_spiral_100` for training or reconstruction, and then use the normalized model and 25 camera poses in `render_mvs_25` for evaluation and comparison.
