Application of few-shot object detection in robotic perception

Abstract

An object detection technique for robotic perception plays a vital role for robots to perform the task that it is functioned to do. In this paper, an efficient and accurate method for object detection for robots is proposed. The paper suggests implementing Few-shot object detection network for robotic vision using the Attention network and attention RPN module. The Multi-relation detector is used to compare two frames and eliminate negative objects from the frame which further enforces the suggested model. Using Contrastive training strategy, the robot is trained to exploit the resemblance between the few-shot support frame and query frame to detect the positive objects and eliminate the negative objects. This method is proposed to help robots perceive the object of interest to perform pick, place, and various other actions. This paper utilizes the COCO dataset to train the network which contains close to 1000 different categories. This method would help accelerate industry 4.0 and has potential in a wide range of applications.