开源连续学习数据集

该数据集包括9504个抽象组合物体模型和190080张高保真渲染图像数据。其中，9504个抽象组合物体模型由正方体、球体、圆柱、圆锥、圆台5种几何基元通过“主物体+副物体”的方式组合而成，涵盖橡胶、金属、玻璃、木质4种材质及红、黄、蓝、绿、紫5种颜色，具备极高的组合自由度与明确的类内多样性；190080张高保真渲染图像数据利用Blender渲染引擎基于光线追踪算法生成，这些图像在环境光照、点光源布局及摄像机视角（每物体20个视角）等方面进行了精细控制，并附带精确的部件级属性标注（Ground-truth），能够支持可解释物体识别及连续学习设定下的新类发现研究。总大小20GB。This dataset comprises 9504 abstract composite object models and 190,080 high-fidelity rendered image samples. The 9504 abstract composite object models are constructed by combining five geometric primitives — cube, sphere, cylinder, cone, and truncated cone — via the "main object + auxiliary object" composition paradigm, covering four material categories (rubber, metal, glass, wood) and five color options (red, yellow, blue, green, purple), and exhibiting extremely high combinatorial freedom and distinct intra-class diversity. The 190,080 high-fidelity rendered images are generated using the Blender rendering engine with ray tracing algorithms. These images feature precise control over environmental lighting, point light source arrangement, and camera viewpoints (20 viewpoints per object), and are accompanied by accurate part-level ground-truth annotations, which can support research on interpretable object recognition and novel class discovery under continual learning settings. The total size of the dataset is 20 GB.

搜集汇总

背景与挑战

背景概述

该数据集包含9504个由几何基元组合而成的抽象物体模型，以及190080张高保真渲染图像，覆盖多种材质和颜色，旨在支持连续学习环境下的可解释物体识别和新类发现研究。数据总量约为20GB，由中国科学院计算技术研究所发布。