欢迎访问《图学学报》 分享到:

图学学报

• 专论:第十九届全国图象图形学学术会议(NCIG2018) • 上一篇    下一篇

基于多角度分割的 360 全景图的显著性检测

  

  1. 1. 北京交通大学计算机与信息技术学院信息所,北京 100044; 
    2. 现代信息科学与网络技术北京市重点实验室,北京 100044
  • 出版日期:2018-12-31 发布日期:2019-02-20
  • 基金资助:
    国家自然科学基金项目(61772066);中央高校基本科研业务费专项资金项目(2018JBM011)

Salient Detection of 360 Panorama Based on Multi - Angle Segmentation

  1. 1. Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China; 
    2. Beijing Key Laboratory of Advanced Information Science and Network, Beijing 100044, China
  • Online:2018-12-31 Published:2019-02-20

摘要: 与常规 2D 图像不同,360 全景图包含当前空间的全部视觉信息,因而在视频监控 和虚拟现实等领域有着广泛的应用,然而用户在某一时刻只能观看到一定的视角,因此,360 全景图的显著性区域检测对于视角预测至关重要。为此,提出了多角度分割的 360 全景图的显 著性检测。首先将全景图进行多角度分割,将分割结果分别投影到立方体上以去除一定畸变; 然后对每个立方体面通过稠密稀疏重建进行显著性计算;最后再将每个面的显著图投影到经纬 映射方式的矩形上,进行多角度融合以获得最终的显著图。通过人工标注的全景图显著区域进 行实验对比,结果表明该算法可以准确检测出 360 全景图的显著区域,并优于当前先进算法。

关键词: 360 全景图, 显著性检测, 多角度分割, 稠密重构误差, 稀疏重构误差

Abstract: Unlike conventional 2D images, 360 panorama contains all the visual information of the current space, so it has a wide range of applications in video surveillance and virtual reality. However, a certain angle is available at a certain time. Therefore, the significant region detection of the 360 panorama is very important to visual angle prediction. To solve this problem, we  propose a multi-angle segmentation based 360 panoramic image saliency detection. Firstly, the panoramic images are cut at multiple angles, and the segmentation results are projected to the cube to remove certain distortion. Then, the salient calculation is conducted for each cube surface through dense and sparse reconstruction. Finally, the saliency images of each surface are projected to the rectangular of the warp and weft mapping, and multi-angle fusion is made to obtain the final salient figure. The results of the 360 panorama test by manual annotation show that the algorithm can accurately detect the saliency and is better than the other methods for the saliency detection of the 360 panorama.

Key words: 360 panorama, saliency detection, multi-angle segmentation, dense reconstruction error, sparse reconstruction error