基于距离谱回归的全景分割方法  

A New Panoptic Segmentation Model Based on Offset Fields

作　　者：孟凡满 丁宇劼 陈帅 罗旺 MENG Fanman;DING Yujie;CHEN Shuai;LUO Wang(School of Information and Communication Engineering,University of Electronic Science and Technology of China,Chengdu 611731 China;School of Electrical and Automation Engineering,Nanjing Normal University,Nanjing 210023 China;Nari Group Corporation/State Grid Electric Power Research Institute,Nanjing 211000 China)

机构地区：[1]电子科技大学信息与通信工程学院,四川成都611731 [2]南京师范大学电气与自动化工程学院,江苏南京210023 [3]南瑞集团有限公司(国网电力科学研究院有限公司),江苏南京211000

出　　处：《西华大学学报（自然科学版）》2020年第4期32-39,共8页Journal of Xihua University:Natural Science Edition

基　　金：国家自然科学基金面上项目(61871087)。

摘　　要：全景分割旨在同时预测图像中每一个像素的语义标签和实例标签,是目前场景解析研究的难点和热点之一。针对现有的全景分割方法采用检测策略提取每个实例区域,无法有效解决实例遮挡的难题,提出一种基于距离谱的深度卷积全景分割模型。不同于现有的基于检测策略的全景分割方法,该模型利用前景像素点与边缘的距离关系构建最大距离和最小距离向量,在存在遮挡的情况下能够有效地刻画出同一实例不同区域的相对关系,从而缓解遮挡问题。为了有效预测最大距离谱和最小距离离谱,提出了一种基于卷积网络的距离谱回归模块。同时,为充分利用距离谱的实例表示优势,设计了2种由距离谱生成实例分割结果的方法。在Cityscapes数据集上进行了大量实验,实验结果表明,全景质量PQ达到了理想效果,本文方法有效。Panoptic segmentation aims to assign both semantic label and instance ID for things class(countable),and only semantic label for stuff class(amorphous),and attracts much interest in recent years.The existing methods focus on detection-then-segment methods.They cannot handle instance overlaps in real scenarios.This leads to insufficient performance.This paper proposes a panoptic segmentation method based on the convolutional neural network.This method models the mutual relationships inside and outside the parted instance via a tailored representation,namely offset fields.This means the maximum and minimum offset vectors correspond to its instance border.Meanwhile,this paper proposes a CNN-based module for predicting those two offsets.Besides,to further explore the potential of the proposed offset fields,we carried out two different manners to obtain the instance result via the offset fields.To verify the effectiveness of the proposed offset fields on alleviating instances overlaps,we conducted extensive experiments on the challenging cityscapes dataset and obtained the good PQ value.

关 键 词：全景分割 距离谱 卷积网络 语义分割 实例分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP751.1[自动化与计算机技术—计算机科学与技术]