SPECSIA-15K
收藏资源简介:
SPECSIA-15K是由韩国科学技术院和中央大学联合创建的面向基于绘图的3D动画新视角增强的风格化数据集。该数据集包含14,980个样本,源自1,498个3DBiCar角色的多视角渲染,每个角色采集10个视角,数据以512×512分辨率的RGBA PNG格式存储,并辅以前景掩膜和位置提示等辅助信号。数据集通过从单视图重建3D模型并重投影至新视角来模拟投影伪影,构建了带伪影的投影与无伪影目标之间的配对监督数据。该数据集旨在解决绘图动画中新视角下轮廓扭曲、纹理缺失等渲染伪影问题,支持训练轻量级后校正模块以提升动画的视角一致性与时序连贯性。
SPECSIA-15K is a stylized dataset jointly developed by the Korea Advanced Institute of Science and Technology (KAIST) and Chung-Ang University for novel view enhancement of drawing-based 3D animation. It contains 14,980 samples derived from multi-view renderings of 1,498 3DBiCar characters, with 10 viewpoints captured for each character. All data is stored in RGBA PNG format at 512×512 resolution, supplemented by auxiliary signals including foreground masks and position prompts. The dataset simulates projection artifacts by reconstructing 3D models from single views and reprojecting them to novel views, thereby constructing paired supervised data between artifact-containing projections and artifact-free target samples. This dataset aims to address rendering artifacts such as outline distortion and texture loss in novel views of drawing-based animations, and supports training lightweight post-correction modules to improve the view consistency and temporal coherence of animations.
数据集概述
SPECSIA-15K 是一个用于绘画驱动3D动画中新颖视角增强的配对风格化数据集。
数据集规模与构成
- 总样本量:14,980 个配对样本(含投影伪影的输入与对应的无伪影目标)
- 角色数量:1,498 个 3DBiCar 角色
- 视角数量:每个角色 10 个均匀采样的视角
- 数据格式:512 × 512 RGBA PNG 图像
- 输入三元组:(Z, Z mask, Z pos) + 真实目标 Y*
数据集划分
| 划分 | 角色数 | 样本数 |
|---|---|---|
| 训练集 | 1,298 | 12,980 |
| 验证集 | 100 | 1,000 |
| 测试集 | 100 | 1,000 |
| 总计 | 1,498 | 14,980 |
数据集构建方法
- 对每个角色渲染10个视角作为无伪影真实目标 (Y*)
- 从前视角重建3D模型
- 投影到相同视角,生成易产生伪影的输入 (Z)
- 每个样本包含三元组 (Z, Z mask, Z pos) 与对应的 Y* 配对
许可证说明
数据集仅包含允许重新分发的渲染图像对,不包含原始 3DBiCar 3D 资产和 Mixamo 动作文件。
下载地址
- Hugging Face 数据集:[https://huggingface.co/(具体链接)]
BibTeX 引用
未公布(标注为 TBA)

- 1SPECSIA: Stylization Dataset for Novel-View Enhancement in Drawing-based 3D Animation韩国科学技术院·电气工程学院; 中央大学·人工智能系 · 2026年




