SnoopyFan/MV4DBench

MV4DBench是一个包含31个项目的多视角4D动画网格基准数据集，用于评估图像/视频到4D网格重建的方法。每个项目从8个方位×N个源帧渲染，并附带一个跟踪的地面真实点云（T=16，N=100k，xyz+法线）用于倒角距离评估。数据集中的项目来源于3D动画的精选子集，包括mixamo角色、动物/对象扫描等。帧数范围从11到757，渲染分辨率为1024×1024，24帧每秒，半径=2.2，仰角=15°。数据集的文件结构包括原始动画网格、元数据文件、不同视角的渲染图像和视频，以及地面真实点云文件。地面真实点云文件在16个均匀间隔的源帧索引处采样，并在[-1,1]^3立方体内归一化。数据集还包含一个UID映射文件，用于记录原始完整名称以便追溯。该数据集被用作ActionMesh（Meta，2024）的多视角测试时间适应扩展的保留评估，具体评估数据可参考相关论文。

MV4DBench is a 31-item multi-view 4D animated mesh benchmark for evaluating image/video → 4D mesh reconstruction methods. Each item is rendered from 8 azimuths × N source frames and ships with a tracked ground-truth point-cloud (T=16, N=100k, xyz+normal) for chamfer distance evaluation. The items are sourced from a curated subset of 3D animations (mixamo characters, animal/object scans, etc.). Frame counts range from 11 to 757; renders are at 1024×1024, 24 fps, radius=2.2, elevation=15°. The dataset layout includes the original animated mesh, metadata files, rendered images and videos from different views, and the ground-truth point cloud file. The ground-truth point cloud is sampled at 16 evenly-spaced source-frame indices and normalized to a [-1,1]^3 cube. The dataset also includes a UID mapping file for traceability of original full names. It was used as held-out evaluation for the multi-view test-time-adaptation extension of ActionMesh (Meta, 2024), with benchmark numbers detailed in the accompanying paper.