Advancing Industrial Vision Research with SyGRID: Synthetically Generated Realistic Industrial Dataset

Abstract Industrial automation depends on accurate object recognition and localization tasks, such as depth estimation, instance segmentation, object detection, and 6D pose estimation. Despite significant advancements, numerous challenges persist, especially within industrial settings. To address these challenges, we propose SyGRID, (Synthetically Generated Realistic Industrial Dataset), a new simulated, realistic dataset specifically designed for industrial use cases. \textcolor{black}{In addition, we propose a robotic vision pipeline based on instance segmentation, 6D pose estimation, and vision-based picking strategies with real-world experiments to prove the effectiveness of the dataset in industrial applications.}Its novelty lies in several aspects: the generated frames are photo-realistic images of objects commonly used in industrial settings, capturing their unique material properties; this includes reflection and refraction under varying environmental light conditions. Moreover, SyGRID includes multi-object and multi-instance cluttered scenes accurately accounting for rigid-body physics. Aiming to narrow the currently existing gap between research and industrial applications, we also provide an exhaustive study on different tasks: namely 2D detection, segmentation, and 6D pose estimation. These tasks of computer vision are essential for the integration of robotic applications such as grasping.SyGRID can significantly contribute to industrial tasks, leading to more reliable robotic operations. By providing this dataset, we aim to accelerate advancements in robotic automation, facilitating the alignment of current progress in computer vision with the practical demands of industrial robotic applications.