MesaTask-10K
收藏资源简介:
MesaTask-10K是一个大规模的桌面场景数据集,包含约10708个由人工精心制作的布局,确保了场景的真实性和物体间复杂的相互关系。该数据集涵盖了6种常见的室内桌子类别,包括办公桌、餐桌、厨房柜台等。MesaTask-10K数据集基于一个包含超过12000个刚性和交互式3D资产的资产库,每个资产都附带详细的语义信息。数据集的创建过程首先由一个预训练的文本到图像模型生成多样化的桌面场景图像,然后通过深度估计、物体检测和3D资产检索构建粗略的3D布局,最后由人工进行细致的布局优化,并通过物理模拟确保物体的碰撞检测。该数据集旨在解决机器人操作任务中任务描述与场景之间的差距,推动基于任务驱动的桌面场景生成研究。
MesaTask-10K is a large-scale desktop scene dataset containing approximately 10,708 manually crafted layouts, which ensures the authenticity of the scenes and the complex inter-object relationships. This dataset covers 6 common indoor table categories, including office desks, dining tables, kitchen countertops, etc. The MesaTask-10K dataset is built upon an asset library containing over 12,000 rigid and interactive 3D assets, with each asset equipped with detailed semantic information. The dataset construction process begins with generating diverse desktop scene images via a pre-trained text-to-image model, then builds rough 3D layouts through depth estimation, object detection and 3D asset retrieval, and finally conducts manual fine-grained layout optimization while ensuring object collision detection via physical simulation. This dataset aims to bridge the gap between task descriptions and scenes in robotic manipulation tasks, and promote research on task-driven desktop scene generation.




