基于多重语义交互的递归式场景理解框架  

Multi-semantic Interaction Based Iterative Scene Understanding Framework

作　　者：姚拓中[1] 左文辉 安鹏[1] 宋加涛[1] YAO Tuo-zhong;ZUO Wen-hui;AN Peng;SONG Jia-tao(School of Electronic and Information Engineering,Ningbo University of Technology,Ningbo,Zhejiang 315016,China;College of Information Science and Electronic Engineering,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]宁波工程学院电子与信息工程学院,浙江宁波315016 [2]浙江大学信息与电子工程学院,杭州310027

出　　处：《计算机科学》2019年第5期228-234,共7页Computer Science

基　　金：国家自然科学青年基金(61502256);浙江省重点研发计划项目(2018C01086);宁波市自然科学基金(2018A610160)资助

摘　　要：传统基于前馈设计的视觉系统已经非常普遍,但其存在的一大缺陷是某个环节出现的错误无法被及时修正,从而影响系统的最终性能。为此,提出了一种简易的交互式框架,其特点在于场景语义的不确定性能够通过不同的视觉分析过程协同工作实现求解和优化。在该框架中,分别使用了3个经典的场景理解算法作为视觉分析模块,不同模块之间利用彼此输出的表面布局、边界、深度、视点和物体类等上下文语义之间的交互以实现各自性能的渐进式提升。提出的方法不需要人为设置约束条件,可根据需求插入新的模块而无须对原有框架和算法进行大的修改,具有良好的可扩展性。基于Geometric Context数据集的实验结果表明,这种基于本征信息交互的反馈式设计通过多次递归后能够有效弥补前馈式系统存在的不足,其中表面布局、边界和视点估计的平均精度提升了5%以上,而物体类的平均检测精度也提升了6%以上,其可成为未来改进视觉系统性能的途径之一。Traditional feed-forward based visual systems have been widely used for years and one fatal defect of this kind of system is that they can’t correct the mistakes by themselves during working,thus resulting in the performance degradation.This paper proposed a simple interactive framework,which solves the semantic uncertainty of the scene through the cooperation of multiple visual analysis processes,leading to scene understanding optimization.In this framework,three classic scene understanding algorithms are used as visual analysis modules and their outputs such as surface layout,boundary,depth,viewpoint and object class are shared for each other by contextual interaction,so as to improve their own performance iteratively.The proposed framework doesn’t need man-made constraints and can add new models in without large modifications of the original framework and algorithms,so it has good scalability.The experimental results on Geometric Context dataset demonstrate that this intrinsic information interaction based system has better flexibility and performs better than traditional feed-forward based systems.The mean accuracy of surface layout,boundary and viewpoint estimation is increased by more than 5%and the mean accuracy of object detection is increased by more than 6%.This attempt can be an efficient way of improving traditional visual systems.

关 键 词：表面布局估计 边界/深度估计 物体/视点检测 多重语义交互 递归式场景理解 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]