室内场景的布局估计与目标区域提取算法

Layout Estimation and Object Region Extraction Algorithm for Indoor Scene

现有的目标提取方法在应用于复杂的室内场景图像时,容易出现小尺寸物体与平面区域中物体被忽视,以及因遮挡造成大物体提取错误等问题。为此,提出一种针对室内RGB-D场景的无监督布局估计与目标区域提取算法。利用3D点云进行平面分割与分类以完成布局估计,采用2种图像分割方法对RGB-D图像做过分割处理,并利用4种相似度衡量方式进行层次分组。在此基础上,根据布局估计的结果,对不同类别的区域采取不同的边界框匹配策略。实验结果表明,该方法无需预训练即可改善目标区域提取效果,在产生较少目标候选区的情况下提高边界框召回率,加快计算速度。

When applying most existing object proposal methods on complex indoor scenes,the results show that there are some problems such as ignoring the small size object and objects in planar regions and detection inaccuracies of big objects caused by occlusion. Aiming at above these problems,this paper proposes a layout estimation and object region extraction algorithm for indoor RGB-D scenes. Firstly,it uses the 3D point cloud for plane segmentation and classification. Secondly,it adopts two segmentation methods using RGB-D data for obtaining crude object segments and then utilizes four similarity measures for hierarchical grouping. Finally,based on the results of layout estimation,it takes diversification strategies to fit bounding boxes for different regions. Experimental result shows that the proposed algorithm can improve extraction efficiency obviously and improve bounding box proposal recall score with fewer object candidates. In addition,it does not need pre-training and has fast calculation speed.