香港科大权龙教授最新视频解读大规模城市场景识别

点击上方 “中国科学信息科学”, 及时获取期刊动态!

《中国科学: 信息科学》(英文版) 第 10 期 MOOP (Multimedia Open Online Paper) 栏目刊登了香港科技大学权龙教授等的论文 “A robust three-stage approach to large-scale urban scene recognition”。 论文介绍了一种稳健的大规模城市场景识别方法。

Presentation

https://v.qq.com/txp/iframe/player.html?vid=q055042ra6p&width=500&height=375&auto=0

Demo

https://v.qq.com/txp/iframe/player.html?vid=u0550b47ppp&width=500&height=375&auto=0

Overview of the paper

Introduction

The three stage approach to large-scale urban scene recognition.

(a) Reconstructed mesh and input images with recovered camera poses.

(b) Joint semantic segmentation.

(c) Object segmentation.

(d) Object abstraction.

Joint semantic segmentation

Building object segmentation

Building abstraction motivation

Contour abstraction

Results and evaluation

Conclusion

• A robust three-stage urban recognition approach able to deal with large scale urban reconstructed models.

• Exploit the multi-view consistency to impose labeling consistency between 2D images and 3D mesh by integrating the constraints in a CRF model.

• The building abstraction algorithm encodes the contextual information of the urban scene structure in the higher-order CRF formation.

复制以下链接访问《中国科学: 信息科学》(英文版) 官方网站免费下载原文:

http://engine.scichina.com/doi/10.1007/s11432-017-9178-8

该文随以下专题出版：

Special Focus on Machine-Type Communications

Science China Information Sciences 

2017, Vol.60, Iss.10

访问以下链接，免费下载专题文章：

http://engine.scichina.com/publisher/scp/journal/SCIS/60/10?slug=Browse