Saliency based proposal refinement in robotic vision

Detecting object grasps from the given image has attracted lots of research concerns in the field of robotic vision. Despite many solutions have been proposed, they tend to simply focus on the detection problem and strongly assume that the object has been placed in the ideal viewing position. In this paper, we propose to refine object proposal based on the saliency measurement. It can be used to refine the object detection results and further guides the self-movement of robotic arm to achieve a better grasping state. First, we dilate the inaccurate proposal to cover more object regions and extract object using saliency-like evaluation measurement. Then, we use superpixel-based sliding windows with various scales and aspect ratios to localize region with highest response. Compared with traditionally exhaustive sliding search, our method reduces the number of sliding windows and hence runs faster. Experiments on public dataset and real test both verify the effectiveness of our proposal method.