RISAT, real-time instance segmentation with adversarial training

Abstract

With the development of artificial intelligence, autonomous driving has gradually attracted attentions from academia and industry. Detecting road conditions correctly and timely is essential to autonomous driving. Thus, we propose a flexible and parallel framework called RISAT for real-time instance segmentation. RISAT improves on YOLOv3 by adding a new parallel branch to generate masks. RISAT can produce a good performance on high-quality segmentation for each instance using GAN. Furthermore, we utilizes ROI class loss on both mask learning for each class and perceptual loss on detailed information. On the benchmark of MS COCO, the frame per second(FPS) of RISAT can achieve 43, which is much faster than that of MNC and FCIS. Besides, the average precision(AP) of RISAT is greater than the previous one-stage object detection method by 0.5.