The paper highlights the importance of large-scale datasets for training models in industrial automation, specifically vision-based robotic grasping. It identifies issues in existing datasets, particularly in the annotation of grasp bounding boxes within the Jacquard Grasp Dataset. The authors propose a Human-In-The-Loop (HIL) method to enhance dataset quality. This method uses deep learning networks to predict object positions and orientations for robotic grasping. Predictions with low Intersection over Union (IOU) values are evaluated by human operators. The refined dataset, named the Jacquard V2 Grasping Dataset, demonstrated significant improvements in training and prediction performance.
Publication date: 9 Feb 2024
Project Page: https://github.com/lqh12345/JacquardV2
Paper: https://arxiv.org/pdf/2402.05747
