Multi-stage Floor Plan Recognition and 3D Reconstruction

The floor plan recognition problem is a sophistical and time-consuming task that requires architects to craft with their professional skills and tools. To solve Raster to vector (R2V) issues, this paper presents a deep-learning-based framework mainly consisting of YOLOv5, deep residual network (DRN), and OCR extraction algorithms. Our work can be interpreted in five steps: 1. Apply YOLOv5 to identify and separate a two-dimensional floor plan main body from the entire figure. Remove the annotation arrows by means of YOLOv5; 3. Extract the main structure data of the floor plan in the light of DRN;4. Apply OCR technology to read the annotation information; 5. Revise the floor plan to a real scale by means of the structured data from DRN recognition. The comparison experiment between our model and the original DRN model is carried out on the classical dataset RJFM (Recognition and 3D Reconstruction of Japan Residential Floor plan Model) and evaluated by a total of four loss functions, i.e. cross-entropy loss, near-neighbor field loss, endpoint regression loss, and multi-task loss. The results show that our model outperforms other methods in 2D floor plan recognition and further isometric 3D reconstruction.