actionbench
收藏资源简介:
ActionBench 是一个用于评估从视频生成动画3D网格的基准数据集,包含128对视频与动画点云的配对样本。每个样本包括:1) 一段16帧的RGBA格式视频(带alpha遮罩);2) 动画点云数据,其形状为(T, V, 6),其中T=16表示关键帧数量,V表示顶点数量(网格表面随机采样的点),6表示每个点的位置(x,y,z)和法向量(nx,ny,nz)。该点云具有跟踪特性,即每个点索引对应随时间变形的相同表面点,提供了密集的时间对应关系。数据集基于ObjaverseXL中的动画对象场景,使用Blender 3.5.1渲染生成。主要评估指标包括:CD-3D(每帧几何精度)、CD-4D(时空一致性)和CD-M(运动保真度)。数据集适用于视频到3D转换、视频分类等任务。
ActionBench is a benchmark dataset for evaluating animated 3D mesh generation from videos, which contains 128 paired samples of videos and animated point clouds. Each sample consists of two parts: 1) a 16-frame RGBA-format video with an alpha mask; 2) animated point cloud data with the shape of (T, V, 6), where T=16 represents the number of keyframes, V denotes the number of vertices (points randomly sampled from the mesh surface), and 6 corresponds to the position (x, y, z) and normal vector (nx, ny, nz) of each point. This point cloud features tracking consistency, meaning each point index corresponds to the same surface point that deforms over time, thus providing dense temporal correspondences. The dataset is generated by rendering animated object scenes sourced from ObjaverseXL using Blender 3.5.1. The primary evaluation metrics include CD-3D (per-frame geometric accuracy), CD-4D (spatiotemporal consistency) and CD-M (motion fidelity). This dataset is suitable for tasks such as video-to-3D conversion and video classification.



