Refining object proposals using structured edge and superpixel contrast in robotic grasping

Abstract

Abstract Grasp detection is an active research branch in robotic field. Most existing works have made strong assumptions, such as the fixed object position and monotonous manipulation background, which facilitate the detection of graspable objects. But the real manipulation condition could be much more complicated. In this work, we propose a novel object perception method. It is able to accurately detect the object, as well as those in cluttered background, and guide the movement of robotic arm to reach a proper grasping state. First, we translate and align the initial proposals according to the structured edge distribution. The aligned proposals have a larger overlap with ground truth at the expense of a little drop in precision. Then, for each superpixel inside the proposal, we use its contrast to high-contrast superpixels and background superpixels, weighted by distance bias, to determine whether it should be included in the refined proposal. Experimental results on both benchmark dataset and robotic task have verified the effectiveness of the proposed method.