Accurate room layout estimation from multi-view panoramas with multi-label graph cut

Estimating room layout from panoramas is a new trend in the holistic reconstruction of the 3D environment. However, a single panorama is easily occluded by walls and furniture, making it hard to reconstruct the whole indoor room accurately and completely. Besides, deep learning room layout estimating methods often perform poorly in unseen scenes. To address this need, this paper proposes an accurate room layout estimation method from multi-view panoramas with multi-label graph cut. The proposed method takes full advantage of each panorama by utilizing multi-label graph cut. First, room layouts of each panorama are estimated with pre-trained deep-learning models and projected to the ground as the labels; then, a geometry-aware ray-casting method is utilized to obtain the initial floorplan; next, the initial floorplan is regularized by multi-label graph cut with the estimated labels from each panorama; in the end, the final layouts of each panorama is obtained by transforming the regularized floorplans and estimated ceiling heights into layouts with panorama geometry. Experiments in the recently released multi-view panoramas dataset show that the proposed method can regularize the initial floorplan to a floorplan with accurate geometry. Furthermore, the accuracy of the layouts surpassed the layout estimation accuracy of the single panorama deep learning models (HorizonNet and LGTNet) and the state-of-the-art self-training layout estimation models with multi-view panoramas by a large margin. • A layout estimation framework considers the geometries and the pre-trained models. • A novel approach effectively fuses multiple layouts with a multi-label graph cut. • Experiments show significant improvements in the accuracy of layout estimation.