FloorSAM: LiDAR Floorplan Reconstruction

• FloorSAM is a framework that reconstructs building floorplans from LiDAR point clouds by merging geometric density maps with zero-shot semantic segmentation.
• The multistage pipeline employs adaptive-resolution projection, image enhancement, and iterative mask filtering to achieve precise room delineation.
• Experimental results demonstrate that FloorSAM surpasses traditional methods with high precision and resilience in noisy, complex indoor layouts.

FloorSAM is a framework for reconstructing building floor plans from LiDAR point cloud data. The method integrates geometric processing of density maps derived from point clouds with the Segment Anything Model (SAM) for zero-shot semantic segmentation, employing a multistage pipeline that fuses semantic and geometric information for robust and accurate floorplan vectorization. FloorSAM addresses limitations in traditional geometry-based and Mask R-CNN-based methods, especially under challenging conditions such as noisy scans, irregular layouts, and incomplete structural information (Ye et al., 19 Sep 2025).

1. Pipeline Architecture

FloorSAM operates through a sequence of seven modules, designed to convert raw LiDAR point clouds into vectorized floorplans:

1. Grid-based Filtering: Initial filtration of the point cloud retains only near-ceiling points to suppress noise from furniture and ground clutter.
2. Adaptive-Resolution Projection: The filtered points are projected into a Bird's-Eye View (BEV) density map at a resolution selected according to input sparsity and extent.
3. Image Enhancement: The density map is enhanced via logarithmic transformation, Gaussian blur, and CLAHE, sharpening room boundaries.
4. SAM-Based Room Segmentation: Adaptive prompt points, chosen from enhanced density peak regions, are fed to SAM for zero-shot room mask generation.
5. Multistage Mask Filtering: Coarse and fine filters remove redundant, merged, or spurious masks, outputting a single mask per room.
6. Joint Mask–Point-Cloud Contour Extraction: Room boundaries are refined using both semantic mask contours and boundary points detected in the point cloud.
7. Contour Regularization and Topology Recovery: Geometric regularization, direction alignment, and topology restoration techniques produce precise, topologically correct polygons, including shared boundaries such as doors.

This staged approach emphasizes the incremental amalgamation of semantic and geometric cues, leading to improved performance over monolithic or single-modality techniques.

2. Key Modules and Algorithmic Details

2.1 Grid-Based Filtering

Given an input point cloud $P = \{ p_i = (x_i, y_i, z_i) \}$, the XY plane is discretized into a uniform grid (cell size $\gamma = 0.1\,$m). For each cell $(m,n)$, only points within $0.1\,$m of the cell’s maximum $z$ are kept:

$$P' = \{ p_j' \mid z_j' \geq \max_{p \in \text{cell}(m,n)} z_p - 0.1 \}$$

This preserves ceiling-adjacent (structural) geometry while filtering out objects and ground-level noise.

2.2 Adaptive-Resolution BEV Projection

For filtered $P'$, BEV density maps $D(m,n)$ are computed at a resolution

$$R = \kappa \cdot \frac{\max(\Delta x, \Delta y)}{\delta}$$

where $\delta$ is the average point spacing and $\gamma = 0.1\,$0 by default. Each point's (X,Y) is mapped to pixel coordinates; density per pixel defines $\gamma = 0.1\,$1. This dynamic resolution ensures detail preservation across varying scale and density.

2.3 Image Enhancement

Enhancement techniques include:

• Log-transform: $\gamma = 0.1\,$2
• Gaussian blur ($\gamma = 0.1\,$3 pixels)
• CLAHE for local contrast optimization

The result is a feature-rich, high-contrast map accentuating room boundaries.

2.4 SAM-Based Room Segmentation

Adaptive Prompt Extraction: Local maxima in $\gamma = 0.1\,$4 are identified (using max-pooling and Gaussian blur). Peaks above threshold $\gamma = 0.1\,$5 become prompt points $\gamma = 0.1\,$6, one or more per room.

Multimask Generation: Each $\gamma = 0.1\,$7 is submitted to SAM for $\gamma = 0.1\,$8 masks ($\gamma = 0.1\,$9–$(m,n)$0), yielding a superset of candidate room masks.

2.5 Multistage Mask Filtering

Coarse Filtering: Masks not meeting criteria for connectivity, area, or deduplication (via IoU thresholding and point-counts) are dropped.

Fine Filtering: Adjacency graphs $(m,n)$1 if $(m,n)$2 identify groups. Inclusion and composite tests prune overlaps and merged-room errors. A greedy set-cover algorithm maximizes coverage while minimizing overlap:

$(m,n)$3

3. Semantic–Geometric Fusion Mechanisms

Fusion occurs at multiple pipeline stages:

• Density Map Generation: Geometry (from $(m,n)$4) is used to derive a raster map, enabling downstream use of image-based models.
• Prompt Selection: Geometric information (density peaks) guides semantic segmentation (SAM prompts).
• Mask Filtering: Both semantic mask properties and geometric statistics (point counts, spatial coverage) inform mask selection.
• Boundary Extraction and Snapping: Final contours fuse semantic mask outlines with geometric boundaries from the point cloud, using a "snapping loss" that aligns polygon vertices $(m,n)$5 to the nearest point-cloud boundary if $(m,n)$6m.

Joint use of semantics and geometry enables robust separation of rooms and precise delineation, particularly when either source alone is ambiguous.

4. Floorplan Polygonization and Topology Recovery

Following mask selection:

• Contour Simplification: Ramer–Douglas–Peucker (RDP) algorithm with an adaptively chosen $(m,n)$7 simplifies mask boundaries to polygons.
• Main-Direction and Orthogonality Alignment: Segments are aligned to dominant room orientation (to nearest $(m,n)$8 or $(m,n)$9 wrt main segment).
• Point-Cloud-Guided Correction: Polygon vertices are snapped to boundary points detected in the geometric point cloud, improving adherence to structural edges.
• Topology Recovery: Adjacent room detection via spatial buffer overlap reconstructs shared elements (e.g., doors), inserting common line segments at doorways into both room polygons to restore floorplan connectivity.

5. Experimental Evaluation

5.1 Datasets and Metrics

• GibLayout: 44 one-floor houses, 3–10 rooms/scene, $0.1\,$0; 12 test scenes with complex geometries.
• ISPRS Indoor Modeling Benchmark: 2 scenes (TUB2, UoM), hundreds of $0.1\,$1; varied sensors.

Metrics:

• Room segmentation: Counts of correctly segmented masks ($0.1\,$295% IoU).
• Floorplan contour: $0.1\,$3, $0.1\,$4: endpoints within $0.1\,$5cm of ground truth.

5.2 Results

GibLayout

Scene	room_true	room_all	room_gt
1	6	6	6
2	7	7	7
3	9	9	10
...	...	...	...
12	10	10	10

Mean $0.1\,$6, $0.1\,$7, derived $0.1\,$8.

ISPRS

Scene	room_true	room_all	room_gt
TUB2	9	9	9
UoM	8	8	8

Mean $0.1\,$9, $z$0, $z$1.

Benchmark Comparison

Method	GibLayout (Prec., Recall)	ISPRS (Prec., Recall)
FloorSP	0.74, 0.77	0.88, 0.86
HEAT	0.72, 0.74	0.73, 0.73
RoomFormer	0.80, 0.82	0.85, 0.87
PolyGraph	0.78, 0.79	0.90, 0.88
PolyRoom	0.76, 0.81	0.89, 0.85
Ours	0.92, 0.94	0.94, 0.95

FloorSAM consistently surpasses all competitors across benchmarks.

5.3 Robustness

• For complex, non-Manhattan layouts, FloorSAM preserves high segmentation accuracy ($z$2) whereas geometric baselines may drop by 30%.
• For noisy scans or incomplete ceilings, only minor recall reductions ($z$3) are observed, attributed to multistage filtering and contour snapping.

6. Codebase and Reproducibility

FloorSAM is accompanied by publicly available code, supplementary materials, and demonstration videos at [https://github.com/Silentbarber/FloorSAM], supporting reproducibility and adaptation to varying LiDAR datasets and building typologies.

7. Significance and Context

By integrating adaptive geometric filtering, density-based BEV mapping, and zero-shot segmentation from SAM, FloorSAM advances the state of the art in floorplan reconstruction. The semantic–geometric fusion mechanism leverages the generalization and zero-shot capabilities of image-based segmentation while maintaining geometric regularity and topological correctness using structural cues from the point cloud. The method does not require task-specific training data or manual annotation for segmentation, providing versatility across diverse indoor environments (Ye et al., 19 Sep 2025).

The approach's resilience to real-world noise, handling of complex layouts, and competitive performance metrics position it as an effective solution for indoor navigation, Building Information Modeling (BIM), and spatial analysis workflows.

References: (Ye et al., 19 Sep 2025)