Deep Multi-Grasp Detection Network v a Augmented Heatmap Regression

Detecting multiple grasping candidates from images simultaneously is more adaptable and efficient in various environments compared to the single grasp detection. In this paper, we establish a novel multi-grasp detection dataset called NMGD, which contains more than 2000 images with up to 14,000 qualified grasping annotations across different objects and layouts. In the scenario of the multi-grasp detection, spatial information inside the images plays a crucial role. However, the fully-connected layer commonly used in the literature for the single grasp detection drops the spatial information and is more likely over-fitting to lose the generalization ability. To avoid the spatial information loss, we propose an augmented heatmap regression method based on Hourglass module which effectively extracts and embeds the spatial information to perform the multi-grasp detection. The proposed method is validated and illustrated by extensive experiments and real robotic grasping trials. Codes and samples of NMGD dataset is made public in https://github.comaingyuYang1997/NMGDN.