Instant and Accurate Instance Segmentation Equipped with Path Aggregation and Attention Gate

Abstract

With the development of GPU and deep learning, there has been great advances in the field of object detection and segmentation. Instance segmentation is one of the most important tasks used in many areas including autonomous vehicles and video surveillance because such areas require both high frames per second (FPS) and high accuracy. In this paper, we propose a method of attaching path aggregation network and attention gate based on real-time instance segmentation model, YOLACT, to increase the accuracy of instance segmentation. As a result of applying the proposed method to the YOLACT framework, the processing speed drops slightly by 2.7%, but the accuracy increases significantly up to 1.4AP, while still maintaining realtime processing of 32.6FPS.