Convolutional multi-grasp detection using grasp path for RGBD images

Generally, most grasp detection models follow the similar frameworks as that in object detection, which use the convolutional neural network to regress the grasp parameters directly. However, grasp detection and object detection are actually different, for the ground truths in object detection are unique while that in grasp detection are not exhaustive. A predicted grasp could still be applicable despite it does not coincide well with ground truth. In this paper, a novel grasp detection model is constructed to make a fairer evaluation on grasp candidate. Instead of using isolated ground truths, the grasp path is introduced to reveal the possible consequent distribution of ground truths. The grasp candidate is first mapped to grasp path, generating the mapped grasp, and the bias between them works as the estimated error for back-propagation. Experiments deployed on grasping dataset as well as real-world scenarios show that our proposed method could improve the detection accuracy. In addition, it can be well-generalized to detect unseen objects.