Understanding of indoor scenes based on projection of spatial rectangles. (September 2018)

• Record Type: Journal Article
• Authors: Wei, Hui; Wang, Luping
• Is Part Of: Pattern recognition. Volume 81(2018:Sep.)
• Page Start: 497
• Page End: 514
• Publication Date: 2018-09
• Subjects: 3D reconstruction -- Quadrangle -- Projected rectangle -- Vanishing point -- Spatial rectangle; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2018.04.017
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A method to efficiently reconstruct 3D indoor scene from a single image is presented. Our method not only can estimate room layout of scene, but also can reconstruct excellent details of scene. The method can cope with clutter without prior training, making it practical and efficient for a navigating robot. The method can estimate locations and positions of spatial rectangles in 3D scenes, without the knowledge of camera's intrinsic parameters, nor of the relation between the camera and world. Abstract: Understanding of indoor scenes has considerable value in mission planning and monitoring in robots. This has become one of the biggest challenges in computer vision because of the diversity and changeability of 3D indoor scenes. Indoor scenes can be considered compositions of many planes in which most common external surfaces are rectangles, such as doors, windows, walls, tables. These spatial rectangles are projected into 2D projections with special geometric configurations, which may enable us to estimate their original orientation and position in 3D scenes. In this paper, the study presents a method to efficiently reconstruct 3D indoor scene without any knowledge of camera's internal calibration. The approach first found quadrangles composed of lines. Through the projection of spatial rectangles, our method not only can estimate room layout of scene, but also can reconstruct excellent details of scene. Due to simple geometric inferences, our method can cope with clutter without prior training, making it practical and efficient for a navigating robot. We compare the room layout estimated by our algorithm against room box ground truth, measuring the percentage of pixels that were classified correctly. Furthermore, we evaluate our ability to fit the indoor scene by comparing against the details that were reconstructed correctly in scene. The experiments showed that our method is capable of reconstructing various structures of indoor environments and that the accuracy and speed of this method meet the requirements a of indoor robot navigation.