Indoor scene understanding based on manhattan and non-manhattan projection of spatial right-angles

Abstract Understanding of indoor scenes has considerable value in computer vision. Most previous methods infer indoor scenes via manhattan assumption. However, attic ceilings do not satisfy manhattan assumption and understanding them remains a big challenge. Non-manhattan ceilings can be seen as compositions of spatial right-angles projections. In this paper, we presented a method to understand indoor scenes including both manhattan structures and non-manhattan attic ceilings from a single image. First, angle projections are detected and assigned to different clusters. Then vanishing points of attic ceilings can be estimated. Third, it is possible to determine the attic ceilings of non-manhattan surfaces. The proposed approach requires no prior training. We compared the estimated attic layout against the ground truth and measured the percentage of pixels that were incorrectly classified. Experimental results showed that the method can understand indoor scenes including both manhattan and non-manhattan attic ceilings, meeting the requirements of robot navigation.