FrameX: An AI System for Intelligent Floorplan Analysis and Applications

Beike has built a comprehensive pipeline for collecting standardized vector floorplan drawings, which retain complete geometric information and enable detailed representation of each layout element.

FrameX, an AI system created by Beike’s Data Intelligence Center, uses these vector floorplans to automatically quantify layout quality, generate descriptive tags, and produce visual interpretations, supporting 18 different business scenarios across the platform.

For individual floorplans, FrameX provides 26 binary tags (e.g., “square layout”, “north‑south ventilation”) and computes three quality dimensions—space scale, lighting, and functional suitability—culminating in an overall composite score, all presented with text, images, and video.

Beyond single‑layout analysis, FrameX quantifies relationships between floorplans by calculating similarity distances, enabling clustering of neighborhoods, districts, and cities; for example, it identified 106,453 distinct floorplan types in Beijing.

The system also offers real‑time similarity search by compressing topological, graphical, and geometric features into vector embeddings, and includes the “拍贝” tool that converts photographed floorplans into standard vectors, currently piloted in Beijing, Chengdu, and Qingdao.

Technically, FrameX follows a layered architecture: a data layer (standard vector database and “拍贝”), a feature layer (interpretation features focusing on points and relational features focusing on overall geometry), and an application layer (tags, scoring, interpretation, similarity queries, and clustering).

Algorithmically, it adapts classic methods such as DFS for ventilation analysis and greedy approaches for squareness, while also employing custom similarity features (absolute/relative layout, simplified/precise contours, area distribution) and large‑scale hierarchical clustering; the “拍贝” pipeline combines deep‑learning heatmaps with integer programming, achieving about 70% accuracy on non‑standard images.

Looking forward, FrameX aims to reduce manual rule‑based constraints, expand AI‑driven interpretation, deepen relationship quantification, and explore renovation and DIY floorplan applications.