TLDR: FloorSAM is a novel framework that uses LiDAR point cloud data and the Segment Anything Model (SAM) to accurately reconstruct building floor plans. It generates density maps from ceiling points, uses SAM for zero-shot room segmentation, and then refines these segments with geometric data to create precise, regularized floor plans, outperforming traditional methods in complex and noisy environments.
Reconstructing accurate floor plans of buildings from 3D point cloud data, often gathered by LiDAR scanners, is a vital technology for many applications. These include indoor navigation systems, Building Information Modeling (BIM), and highly precise indoor measurement. However, traditional methods for this task have faced significant hurdles. They often struggle with noise in the data, have limited ability to adapt to new environments, and can lose important geometric details, which severely impacts the accuracy of measurements.
Introducing FloorSAM: A Novel Approach
To overcome these challenges, researchers have developed an innovative framework called FloorSAM. This system integrates room-height point cloud density maps with the powerful guided segmentation capabilities of the Segment Anything Model (SAM). The goal of FloorSAM is to significantly enhance the precision of floor plan reconstruction from LiDAR point cloud data.
FloorSAM works by first applying a grid-based filtering technique to retain elevation point clouds that are close to the ceiling in each region. This, combined with adaptive resolution projection and image enhancement methods, generates a top-down density map. This map is crucial for improving the robustness and accuracy of spatial feature measurements.
How FloorSAM Works: A Step-by-Step Breakdown
The FloorSAM framework leverages SAM’s zero-shot learning ability to achieve high-fidelity room segmentation. This means it can segment rooms accurately without needing to be specifically trained on a large dataset of annotated floor plans, making it highly adaptable to diverse building layouts. After this initial segmentation, high-quality room masks are generated using adaptive prompt points, followed by a multi-stage filtering process to extract precise semantic masks for individual rooms.
A key innovation of FloorSAM is its joint analysis of both the mask (semantic) and point cloud (geometric) data. This allows for accurate contour extraction and regularization, integrating semantic segmentation with geometric information to produce highly accurate room floor plans and recover the topological relationships between rooms (e.g., how rooms connect through doorways).
The entire process can be broken down into three main steps:
1. Preprocessing and SAM Integration: The system starts by generating a density map from point clouds near the ceiling, after some initial noise reduction. Based on this map, numerous prompt points are created. Both the density map and these prompt points are fed into SAM for comprehensive segmentation, producing many potential room masks.
2. High-Quality Mask Filtering: Since SAM can generate redundant or imperfect masks, a two-step filtering process (coarse-to-fine) is applied. This stage selects only the highest quality single-room masks that accurately represent the semantics of each room, using geometric constraints and semantic validation.
3. Joint Contour Extraction and Topology Recovery: Finally, the refined room semantics and their corresponding projected point clouds are used together to draw regularized contours. This step also detects connection segments to recover the topological relationships between rooms, effectively mapping out the entire floor plan.
Also Read:
- Advancing 2D-3D Sensor Alignment for Autonomous Systems with CrossI2P
- Achieving Clearer Object Edges in AI Depth Perception
Performance and Advantages
Experiments conducted on public datasets like GibLayout and ISPRS, as well as self-collected data, have validated the effectiveness of FloorSAM. The results show significant improvements in measurement accuracy, recall, and overall robustness compared to traditional approaches, especially in noisy environments and complex layouts. FloorSAM consistently achieves average precision and recall rates above 0.90 on these datasets.
Compared to other methods like Mask R-CNN, FloorSP, HEAT, RoomFormer, PolyGraph, and PolyRoom, FloorSAM demonstrates superior performance in both room semantic segmentation and floor plan reconstruction. It produces more complete and accurate masks and floor plans, even in challenging scenarios with non-Manhattan structures (i.e., buildings not strictly adhering to right angles).
While FloorSAM represents a significant leap forward, the researchers acknowledge that the method still faces challenges when point clouds lack sufficient ceiling data. Future work aims to incorporate ground semantics to further improve segmentation generalization and utilize image-based segmentation to recover structures like windows.
For more technical details, you can refer to the full research paper: FloorSAM: SAM-Guided Floorplan Reconstruction with Semantic-Geometric Fusion.
