Statistical Reconstruction of Indoor Scenes

In this paper we consider the problem of processing scanned datasets of man-made scenes such as building interiors and office environments. Such datasets are produced in huge quantity and often share a simple structure with sharp crease lines. However, their usual acquisition with mobile devices often leads to poor data quality and established reconstruction methods fail – at least at reconstructing sharp features. We propose to overcome the lack of reliable information by using a strong shape prior in the reconstruction method: we assume that the scene can be represented as a collection of cuboid shapes, each covering a subset of the data. The optimal configuration of cuboids is found by formulating the reconstruction problem as a discrete maximum a posteriori (MAP) optimization in a statistical sense. We propose a greedy algorithm which iteratively extracts new shape candidates and optimizes over the shape of the cuboids. A new candidate is selected by scoring its ability to reconstruct previously uncovered data points. The iteration converges at the first significant drop in the score of new candidates. Our method is fast and extremely robust to noisy and incomplete data which we show by applying it to scanned datasets acquired with different devices.