Structural Line Detection for Room Layout Estimation from a Painting

Immersive experiences that combine visual and auditory elements can enhance the appreciation of art. One such approach involves 3D visualizations created by digitizing various forms of art. This study focuses on generating 3D representions of interior paintings, particularly those with geometric relationships such as parallel lines, from a 2D image. To generate threedimensional images from interior paintings, estimating room layouts by segmenting walls, ceilings, and floors is highly effective. While existing models are trained on photorealistic images, their application to paintings is challenging due to unclear outlines, unrealistic lighting, and shadows inherent to artistic styles. Additionally, no database currently exists for paintings with corresponding room layout annotations. In this paper, we propose a confidence-based approach to detect structural lines in painting images. This method enables the generation of 3D room layouts from paintings using photorealistic existing datasets. The effectiveness of our approach is demonstrated through both quantitative and qualitative experiments.