Data-Link

Data-Link是一个专为工业环境下模型到真实世界转移设计的高保真数据集，由爱尔兰IMR的工业指导委员会支持开发。该数据集包含60个制造相关的3D对象类别，总计3000个训练图像和1800个测试图像，每个图像分辨率为640×480像素。数据集通过使用高分辨率相机捕捉3D对象，并利用定制的神经网络DiscNet提取对象特征，包括纹理、材料、形状和惯性动态。数据集的创建过程涉及将这些特征转移到数字双环境中，并通过域随机化技术进行标注，以增强模型的泛化能力。Data-Link数据集主要用于解决机器人抓取和操作中的6D姿态估计问题，通过提供高保真的模拟数据，帮助模型更好地适应真实世界环境，减少模拟与现实之间的差距。

Data-Link is a high-fidelity dataset specifically designed for model-to-real-world transfer in industrial environments, developed with support from the Industrial Steering Committee of Ireland’s IMR. This dataset includes 60 manufacturing-related 3D object categories, with a total of 3000 training images and 1800 test images, each having a resolution of 640×480 pixels. High-resolution cameras are utilized to capture the 3D objects, and a custom neural network named DiscNet is applied to extract object features including texture, material, shape, and inertial dynamics. The dataset creation workflow involves transferring these features into a digital twin environment, and conducting annotation via domain randomization techniques to enhance the model's generalization capability. The Data-Link dataset is primarily intended to address the 6D pose estimation task in robotic grasping and manipulation. By providing high-fidelity simulated data, it assists models in better adapting to real-world environments and narrowing the sim-to-real gap.