Performance Analysis of Point Cloud Edge Detection for Architectural Component Recognition

Abstract

With the advancement of 3D sensing technologies, point clouds have become a key data format in the construction industry, supporting tasks such as as-built verification and BIM integration. However, robust and accurate edge detection from unstructured point cloud data remains a critical challenge, particularly in architectural environments characterized by structured geometry and variable noise conditions. This study presents a comparative evaluation of two classical edge detection algorithms—Random Sample Consensus (RANSAC) and Density-Based Spatial Clustering of Applications with Noise (DBSCAN)—applied to terrestrial laser-scanned point cloud data of eight rectangular structural columns. After preprocessing with the Statistical Outlier Removal (SOR) algorithm, the algorithms were evaluated using four performance criteria: edge detection quality, BIM-based geometric accuracy (via Cloud-to-Cloud distance), robustness to noise, and density-based performance. Results show that RANSAC consistently achieved higher geometric fidelity and stable detection across varying conditions, while DBSCAN showed greater resilience to residual noise and flexibility under low-density scenarios. Although DBSCAN occasionally outperformed RANSAC in local accuracy, it tended to over-segment edges in high-density regions. These findings underscore the importance of selecting algorithms based on data characteristics and project goals. This study establishes a reproducible framework for classical edge detection in architectural point cloud processing and supports future integration with BIM-based quality control systems.

1. Introduction

With the rapid advancement of 3D sensing and spatial data acquisition technologies, applications in fields such as architecture, construction, robotic vision, and autonomous systems have expanded significantly [1,2,3]. Among these, point clouds have emerged as a core data format in the construction industry, supporting diverse tasks such as structural analysis, progress monitoring, and Building Information Modeling (BIM) integration [4,5,6,7]. Point clouds represent unstructured raw spatial data, whereas 3D models—such as BIM—are structured representations that incorporate both geometric and semantic information. This distinction is crucial, as edge features extracted from point clouds often serve as intermediaries in converting raw scans into semantically rich models suitable for design, analysis, or verification tasks [8]. Unlike traditional 2D images, point clouds provide three-dimensional spatial data with fine geometric detail, enabling precise representation of real-world objects and environments.

In the context of 3D models, an edge refers to a significant geometric boundary or discontinuity, often indicating transitions between planar or curved surfaces. These edges are critical for recognizing object boundaries, reconstructing topology, and facilitating tasks such as reverse engineering, object recognition, and automated modeling workflows [9]. Edge detection in 3D point clouds plays a critical role in numerous downstream applications, including core geometry extraction, object segmentation, and 3D model reconstruction for as-built verification and BIM enrichment [10]. However, edge detection in 3D data is fundamentally different from its 2D counterpart. Traditional filters such as Canny, Sobel, or Laplacian operate on structured grids, whereas 3D point clouds are irregular and unstructured, lacking inherent neighborhood definitions. To overcome these limitations, various 3D edge detection methods have been developed, including geometry-based approaches like RANSAC, density-based clustering such as DBSCAN, as well as graph-based and learning-based techniques. Each approach differs in its handling of noise, surface irregularities, and computational complexity [9]. Moreover, point clouds often contain noise, missing data, varying density, and sometimes encode reflectance values that must be considered.

To address these challenges, various 3D edge detection techniques have been developed across non-architectural domains such as robotics, autonomous driving, agriculture, and manufacturing. Han et al. [11] introduced an octree-based method using normal vectors and geometric indicators to improve edge preservation. Xia et al. [12] proposed a Graph Snapping technique to detect road structures from mobile LiDAR scans more efficiently. Ahmed et al. [13] developed an edge and corner detection algorithm based on local symmetry and adaptive density thresholds for robotic welding. Wang et al. [14] employed a modified RANSAC algorithm for road edge extraction with a threshold-based noise filter, while He et al. [15] applied RANSAC to farmland boundary detection with enhanced angular accuracy. Chen et al. [16] utilized DBSCAN for lane boundary clustering under noisy conditions, and Guo et al. [17] demonstrated DBSCAN’s effectiveness in segmenting complex transportation scenes with variable point densities.

These studies demonstrate robustness in their respective domains, but many lack validation in architectural settings, where edge semantics, occlusion, and geometric regularity differ significantly. Despite progress in general-purpose methods, relatively few studies have addressed the specific requirements of architectural point clouds, which often involve structured and rectilinear elements such as walls, columns, and floors. Moreover, real-world architectural data typically includes various sources of noise, occlusion, and partial scans, requiring algorithms with both geometric precision and robustness to irregularities.

In evaluating the quality of architectural geometry representation, this study adopts three quantitative indicators—linearity, Mean Deviation Error (MDE), and edge detection accuracy—as previously applied in architectural point cloud studies [6,9,18,19,20], which collectively assess geometric regularity, dimensional fidelity, and the correct identification of structural boundaries.

This study aims to evaluate the performance of two representative algorithms—RANSAC and DBSCAN—in identifying architectural components, specifically rectangular structural columns, from terrestrial laser-scanned point cloud data. The reference BIM model was constructed at Level of Detail (LOD) 300 to ensure sufficient geometric granularity for reliable edge-based comparison with the scanned data. Columns were selected as the target element due to their ubiquity, structural importance, and controlled geometric form, making them suitable for foundational benchmarking. While the current scope focuses on rectangular forms, the methodology may be extended to more diverse architectural geometries in future studies.

Among the diverse families of 3D edge detection methods, RANSAC and DBSCAN were deliberately selected for benchmarking due to their complementary characteristics in indoor architectural contexts. RANSAC represents a geometry-based fitting approach well-suited for structured environments, while DBSCAN exemplifies a density-based clustering approach resilient to irregularities and noise. By comparing these two distinct methods under identical conditions, this study establishes a representative performance baseline for edge detection in architectural point clouds. Ultimately, this research supports automated quality verification and BIM integration in digital construction workflows by establishing a reliable performance baseline for classical edge detection algorithms.

2. Related Work

Early model fitting techniques such as Random Sample Consensus (RANSAC) laid the foundation for many point cloud processing algorithms. Originally proposed by Fischler and Bolles [21] in the context of image analysis and automated cartography, RANSAC offers a robust framework for fitting models to data contaminated with a high proportion of outliers. Its iterative process of minimal sample selection and consensus evaluation has since been widely adopted in 3D point cloud analysis—particularly for edge and feature detection tasks where measurement noise and structural irregularities are common. The core principle of isolating inlier subsets has enabled diverse applications, ranging from primitive shape fitting to architectural feature extraction. However, in indoor architectural environments, repetitive rectilinear geometries and partial occlusions often reduce RANSAC’s robustness, requiring careful parameterization and validation.

Other early methods such as Region Growing also provided essential tools for point cloud segmentation. These approaches rely on smoothness constraints and normal vector continuity to group adjacent points into meaningful regions. Rabbani et al. [19] proposed a robust implementation of this method, demonstrating its effectiveness in separating curved and planar features from noisy data. Accurate estimation of surface normal is critical in such processes. To this end, Mitra and Nguyen [22] introduced a least-squares fitting approach that accounts for neighborhood size, curvature, and point density—factors essential for reliable geometric interpretation. Yet, in building interiors where clutter and occluding objects disrupt smooth surface continuity, Region Growing can struggle to delineate structural edges.

Hough Transform has also been applied for detecting parametric primitives such as lines, circles, and planes in 3D scans, leveraging its tolerance to noise and missing data [23]. In particular, it has been used for building model reconstruction by extracting roof and façade boundaries, demonstrating applicability to architectural domains. However, its computational cost and limited adaptability to large-scale indoor scans restrict its broader use in architectural environments.

While such foundational techniques remain influential, their application to architectural environments introduces new challenges. The repetitive geometry, occlusions, and partial visibility typical of building interiors demand more specialized approaches, particularly those that can align with semantic-rich BIM models. Tang et al. [6] provided a comprehensive review of as-built BIM workflows and highlighted the importance of geometric primitives such as edges and planes in this context. Duanmu et al. [24] further proposed a preprocessing algorithm (ANPDA) to enhance edge clarity in noisy indoor data, showing adaptability from forestry to architectural applications.

Recent advances have explored learning-based approaches to overcome these limitations. Pan et al. [25] applied graph convolutional networks with edge-aware features to extract detailed building elements, though the method was computationally intensive and sensitive to occlusion. Similarly, Ochmann et al. [20] developed a system to reconstruct parametric indoor models, implicitly leveraging edge cues to define geometric boundaries between structural components. Super Point and LoFTR have also demonstrated strong robustness in extracting edge-like features from noisy or low-texture data [26,27]. However, their reliance on large, labeled datasets and high computational cost limits their current utility in architectural-scale scans.

In addition to these individual algorithms, hybrid techniques have gained interest. Scitovski et al. [28] proposed a method that combines RANSAC and DBSCAN for detecting multiple geometric shapes in complex scenes. Although not tailored for buildings, their approach illustrates the potential synergy between model fitting and density-based clustering—an idea adopted in this study.

Taken together, these studies showcase both the evolution and limitations of edge detection in architectural point clouds. While recent methods exhibit promising accuracy, many remain difficult to generalize, scale, or validate against BIM-based ground truth. This highlights the need for foundational studies that systematically evaluate classical algorithms using real-world architectural data. To address this, the present study selects RANSAC and DBSCAN—two widely used but fundamentally distinct approaches—for comparative benchmarking of edge detection performance in indoor building environments.

3. Research Methodology

3.1. Scope and Approach

The objective of this study is to evaluate the performance of RANSAC and DBSCAN algorithms in recognizing architectural elements from indoor point cloud data. To this end, point clouds were acquired from an educational facility using a laser scanner, and edge detection was performed after applying a noise removal preprocessing step.

Although pillars exhibit relatively simple geometry, their contours can vary slightly depending on construction quality or structural defects. These subtle differences are effectively captured through edge-based analysis, making pillars a suitable subject for evaluating detection algorithms.

Eight rectangular pillars located in classrooms and hallways were selected to reflect diverse indoor conditions, such as variations in pillar location, number of exposed faces, and proximity to reflective or obstructive elements. Figure 1 illustrates the spatial layout and contextual placement of these pillars, providing a visual overview of the spatial diversity considered in the experiment. This diversity ensures that the edge detection algorithms are tested under realistic conditions, enhancing the generalizability of the results.

Room A and Room B include pillars that are adjacent to walls, where three faces of each pillar can be captured. Room C contains pillars located at the edge of space, exposing only two faces. In Hallway H, the rear of the pillar faces a window, presenting an environment where noise may occur due to light reflection.

This study takes these conditions into account to perform a comparative analysis of the edge detection performance of the RANSAC and DBSCAN algorithms on point cloud data.

The overall research process is illustrated in Figure 2. This study is conducted based on a review of prior research. It is structured to perform edge detection using RANSAC and DBSCAN algorithms after applying a noise removal preprocessing step, followed by a comparative performance analysis. The comparison criteria include edge detection quality, robustness to noise, and density-based performance. Detailed descriptions of the equipment used, and data processing methods applied in this study are provided in Section 3.2.

3.2. Point Cloud Acquisition and Preprocessing

To ensure accurate and consistent data collection, this study employed the Leica BLK360 laser scanner as the sole device for acquiring all point cloud data. The device offers a scan speed of 360,000 pts/sec and a scanning range of 360° horizontally and 300° vertically. Its measurement accuracy is 4 mm at 10 m, and the maximum measurable distance is 60 m, making it suitable for indoor spatial analysis. A total of eight pillars across four architectural spaces (classrooms and corridors) were scanned, with each pillar captured from four scanning positions. The scanner was mounted on a tripod at a height of approximately 1.2 m.

The collected scans were registered using Leica Cyclone FIELD 360. Subsequently, Cyclone 3DR was used to define spatial boundaries and extract only the pillar regions (Figure 3). To minimize detection interference from adjacent objects, data were selected with a 200 mm margin around each pillar.

To enhance data quality and remove noise, the Statistical Outlier Removal (SOR) algorithm was employed. The SOR algorithm identifies the k-nearest neighbors (KNN) for each point and calculates the mean distance among them; points whose mean distance exceeds a statistical threshold are classified and removed as outliers. Because the algorithm automatically detects and filters noise based on the mean and standard deviation, it enables quantitative noise removal.

The SOR parameters were carefully adjusted to balance noise elimination with data preservation, ensuring that essential point cloud information for edge detection was retained. Specifically, the number of neighbors (nb_neighbors) was set to 20, and the standard deviation ratio (std_ratio) was set to 2.0. The preprocessing algorithm was implemented in a Python 3.7.9 environment.

Table 1. Comparison of data before and after preprocessing (application of SOR algorithm).

	Original Point Cloud (Before SOR)	Processed Point Cloud (After SOR)	Removed Noise Points	Removal Rate (%)
A-1	140,151	138,367	1784	1.27
A-2	184,077	181,306	2771	1.51
B-1	204,001	201,005	2996	1.47
B-2	208,644	205,505	3139	1.50
C-1	401,229	391,489	9740	2.43
C-2	494,001	478,668	15,333	3.10
H-1	178,264	176,277	1987	1.11
H-2	202,914	200,900	2014	0.99

summarizes the effects of the SOR preprocessing algorithm on the eight pillar datasets. A small percentage of noise points (approximately 1–3%) was removed across all cases. The denoised point clouds were then used as input for the subsequent RANSAC and DBSCAN analysis.

3.3. Edge Detection Algorithm

3.3.1. RANSAC Algorithm

The RANSAC (RANdom SAmple Consensus) algorithm, first introduced by Fischler and Bolles [21], is a model-based method designed to identify specific geometric structures, such as lines or planes, within point cloud data. It operates by iteratively selecting random subsets of the input data to generate candidate models and evaluates them based on how well they fit the remaining data.

Formally, consider a dataset $D$ and a parametric model M($\theta$) where $\theta$ represents the model parameters. In each iteration, a minimal sample set S ⊂ $D$ is randomly selected to estimate $\theta$ (1). The model is then fitted to this subset, and the consensus set is determined by where $\delta$ is a predefined threshold indicating how close a data point must be to the model to be considered an inlier.

$$C\left(\theta \right)=\{x\in D|error(x,\theta )<\delta \}$$

(1)

The algorithm is repeated for a fixed number of iterations or until a model with a sufficiently large consensus set is found. The model with the largest number of inliers is then selected as the final estimate. This iterative random sampling strategy allows RANSAC to effectively avoid the influence of outliers. RANSAC has been widely applied in point cloud analysis tasks, including edge and boundary detection in 2D and 3D environments [14,15]. Recent studies have also explored combining RANSAC with clustering techniques such as DBSCAN to improve robustness in complex geometries [28].

The application process of the RANSAC algorithm for edge detection is as follows. First, a subset of points is randomly selected from the point cloud to establish an initial model. A linear or curved geometric model is then fitted to the selected samples. Next, all points in the dataset are evaluated: those that fall within a predefined error threshold are classified as inliers, while those that deviate significantly are treated as outliers. This process is repeated multiple times, and the model that accumulates the highest number of inliers is chosen as the optimal edge model.

RANSAC inherently includes an outlier rejection mechanism, making it particularly useful in environments with clearly defined geometric structures such as walls, pillars, and floors. Moreover, its iterative model optimization enables high detection accuracy. However, its performance is sensitive to the configuration of parameters such as the sampling threshold and the number of iterations, which can influence both computational efficiency and detection precision.

In this study, the RANSAC algorithm was implemented in a Python-based environment using preprocessed point cloud data from eight pillars. The distance threshold was set to 0.013 m so that only points within this range were considered as edge candidates. The number of iterations was fixed at 500, and the minimum number of inliers required was set to 200 to prevent unreliable line detection due to insufficient inliers. Parameter values were selected within ranges commonly adopted in prior RANSAC-based geometric extraction studies, such as Yang et al. [29], to ensure stable and accurate edge detection results across all samples.

3.3.2. DBSCAN Algorithm

The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is a density-based clustering algorithm that analyzes local variations in point density to form clusters. Originally proposed by Ester et al., DBSCAN has proven effective in various spatial data applications, including 3D point cloud clustering [30]. Its ability to form arbitrarily shaped clusters and identify noise makes it especially suitable for edge detection in architectural and engineering datasets [28]. Unlike traditional partition-based clustering methods, DBSCAN does not require the number of clusters to be predefined. Instead, it automatically forms clusters based on data density, making it effective for detecting arbitrarily shaped clusters and identifying noise points. Mathematically, DBSCAN defines a neighborhood around a data point $\mathcal{x}\in {\mathcal{R}}^{\mathcal{d}}$ using a radius parameter $\epsilon$ (2). The $\epsilon$-neighborhood of $\mathcal{x}$ is given by:

$${\mathsf{{\rm N}}}_{\epsilon \left(\mathcal{x}\right)}=\left\{\mathcal{y}\in D |‖\mathcal{x}-\mathcal{y}‖\le \epsilon \right\}$$

(2)

A point x is classified as a core point if the number of points within its $\epsilon$-neighborhood satisfies the threshold MinPts, which is a user-defined value representing the minimum number of points required to form a dense region (3).

$$\left|{\mathsf{{\rm N}}}_{\epsilon \left(\mathcal{x}\right)}\right|\ge \mathrm{MinPts}$$

(3)

A point y that lies within the ε-neighborhood of a core point $\mathcal{x}$ is considered directly density-reachable from $\mathcal{x}$. DBSCAN recursively connects density-reachable points to form clusters, and points that do not belong to any cluster are classified as noise [6].

The application of the DBSCAN algorithm to edge detection proceeds as follows: first, the local density of the point cloud data is analyzed to distinguish between high-density regions (edges) and low-density regions (non-edges). Then, points in the high-density regions are connected to form clusters, and points that do not satisfy the minimum number of points (MinPts) within the distance threshold (eps) are treated as outliers and removed. Since DBSCAN does not assume any specific geometric shape, it is particularly advantageous for detecting curved or non-uniform edges.

Owing to its ability to adaptively detect dense regions, DBSCAN is particularly effective for analyzing complex point cloud structures. In construction-related applications, it has also been utilized for preprocessing and segmentation tasks, particularly for isolating dense geometric features from noise [24].

In this study, the DBSCAN algorithm was applied under the same conditions and datasets as the RANSAC algorithm. The experiment used eight preprocessed pillar point cloud datasets in a Python environment. The distance threshold (eps) was set to 0.015 m to include candidate points within this range for clustering, and the minimum number of points (MinPts) was set to 30 to ensure that only points meeting the minimum density requirement were classified as part of a cluster. Parameter values were selected based on both empirical testing and established guidelines in prior DBSCAN research. Specifically, eps was set to 0.015 m and MinPts to 30, following heuristic norms related to data dimensionality and k-distance analysis. This approach is consistent with studies highlighting the sensitivity of DBSCAN performance to eps selection and the use of nearest-neighbor–based strategies for parameter tuning [31]. Such alignment with established practices ensures that the chosen configuration effectively detects dense edge clusters while minimizing noise sensitivity.

3.4. Selection of Comparison Criteria

3.4.1. Internal Quality Indicators (Edge Geometry-Based)

To accurately evaluate edge detection algorithms, a systematic comparison based on diverse performance metrics is essential. Accordingly, this study selected key evaluation criteria that comprehensively assess the performance of each method under varying conditions, reflecting the diverse challenges commonly encountered in real-world architectural environments, such as varying surface reflectivity, partial occlusions, and irregular object geometries. The selected criteria include edge detection quality analysis, robustness to noise, and density-based performance analysis. These factors are critical for determining the practical applicability of an algorithm when processing point cloud data from architectural environments.

In particular, the study assessed algorithmic performance using the number of detected edges and the average linearity of edge clusters. Average linearity is calculated based on the eigenvalue ratio derived from principal component analysis (PCA) of each detected edge cluster, reflecting the linearity and geometric consistency of the detected features.

Mathematically, for each detected edge cluster $\mathcal{i}$, let λ₁ ≥ λ₂ ≥ λ₃ be the PCA eigenvalues. The linearity $L_{\mathcal{i}}$ is calculated as:

$$L_{\mathcal{i}}={\displaystyle \frac{{\lambda }_1-{\lambda }_2}{{\lambda }_1}}$$

(4)

The average linearity over all $\mathrm{N}$ detected edge clusters is then given by:

$$L_{avg}= {\displaystyle \frac{1}{N}}\sum _{\mathcal{i}=1}^N{L}_{\mathcal{i}}$$

(5)

where $\mathrm{N}$ is the total number of detected edge clusters. This formulation ensures ${\mathcal{L}}_{\mathcal{i}}$ approaches 1 for perfectly linear clusters and decreases for less linear shapes.

These indicators enable consistent and objective comparisons even in the absence of manually defined ground truth data, making them especially suitable for architectural applications. Furthermore, since average linearity does not rely on annotated ground truth, it allows for reproducible and consistent evaluations across different algorithms and datasets. These metrics effectively capture the algorithm’s sensitivity to structural boundaries and its ability to preserve geometric coherence. By focusing on measurable edge characteristics rather than manual labels, this approach enhances the objectivity and reproducibility of the evaluation.

In prior research, the evaluation of edge detection algorithms has largely focused on accuracy metrics. For example, Ni et al. (2016) [9] assessed edge and feature line detection performance using precision and false detection rate based on reference data, with a strong emphasis on geometric accuracy. However, this study extends beyond traditional accuracy-based evaluation by introducing two additional criteria: robustness to noise and density-based performance. While these criteria are still related to detection accuracy, they were specifically introduced to assess the robustness and spatial consistency of algorithms in response to the complex realities of architectural environments. Such multifaceted evaluation is essential, as architectural point cloud data often exhibit irregularities, occlusions, and varying densities that cannot be fully captured by accuracy alone.

The robustness to noise evaluation measures the influence of residual noise—remaining even after preprocessing—on edge detection performance. This metric is crucial for understanding how well an algorithm can maintain performance in realistic, imperfect data conditions. Density-based performance analysis, on the other hand, evaluates the impact of point density on detection accuracy and edge detection effectiveness, particularly in cases where scanning conditions cause uneven data distribution.

By introducing these two novel evaluation criteria and conducting a structured analysis, this study aims to clearly identify the strengths and limitations of each algorithm. This comprehensive approach significantly contributes to establishing optimal application strategies for edge detection algorithms in the processing of architectural point cloud data. The subsequent sections provide detailed analyses of the methodologies and results based on each evaluation criterion.

3.4.2. External Accuracy Indicators

To ensure an objective and geometrically meaningful evaluation of edge detection accuracy, this study adopts a Building Information Modeling (BIM) model as a design-based reference proxy rather than an absolute ground truth. The BIM model was developed through a CAD-to-BIM conversion process based on architectural floor plans of the space containing the target columns. While it represents the intended design geometry, including column positions, heights, cross-sectional shapes, and spatial orientations, potential deviations from the as-built conditions are acknowledged. To enhance the accuracy of the BIM model, actual field measurements were conducted using a laser distance meter. For example, the H-1 column was designed with a height of 2.700 m, but the measured height was 2.669 m; likewise, the designed width of 0.800 m was measured as 0.823 m, indicating small discrepancies. These measured deviations were incorporated into the BIM model wherever possible to improve its suitability as a comparative baseline. This process ensured that the reference model closely reflected the actual site conditions, minimizing bias in the subsequent geometric evaluation.

For spatial alignment, a two-step registration procedure was performed between the detected edge point clouds and the BIM geometry. First, geometric features were extracted from both datasets to achieve initial alignment, followed by refinement using the Iterative Closest Point (ICP) algorithm. This hybrid registration approach improves robustness against noise and partial data, ensuring high-precision matching of edge points to the BIM surface. To validate the reliability of this registration, the quality of the ICP results was quantitatively assessed. The final ICP refinement achieved an average RMSE of approximately 0.003 m and a mean residual alignment error of 0.002 m, values which are within the nominal accuracy tolerance of the BLK360 scanner (4 mm at 10 m). These results confirm that the alignment was sufficiently precise, ensuring that the subsequent deviation metrics (e.g., C2C and RMSE) primarily reflect geometric discrepancies between the as-built pillars and the BIM reference rather than misregistration artifacts.

The evaluation metric used is the Cloud-to-Cloud (C2C) distance, which computes the average Euclidean distance from each detected edge point to its nearest surface on the BIM model. This means the mean C2C value represents the typical geometric deviation, while the standard deviation (STD) quantifies the spatial consistency of alignment across different regions. In addition, the root mean square error (RMSE) is also reported to capture the influence of large local deviations, which can arise from partial occlusions, surface irregularities, or subtle mismatches between the as-built geometry and the BIM reference. Together, these indicators provide a continuous and interpretable measure of how precisely the extracted edge geometries conform to the reference BIM.

In addition, the edge detection accuracy (%) is calculated as:

$$Accuracy\left(\%\right)={\displaystyle \frac{N_{correct}}{N_{total}}}\times 100$$

(6)

where $N_{total}$ is the total number of detected edge points, and $N_{correct}$ is the number of detected edge points whose C2C distance to the BIM reference edge is within the positional tolerance τ.

Furthermore, the Mean Deviation Error (MDE) is defined as:

$$MDE= {\displaystyle \frac{1}{M}}\sum _{j=1}^M\left|d_j\right|$$

(7)

where $d_j$ is the C2C distance between the j-th detected edge point and its nearest point on the BIM reference edge, and M is the total number of matched points.

These C2 C-based metrics were applied to edge detection results obtained from both RANSAC and DBSCAN to quantitatively evaluate their geometric performance. Visual comparisons and analysis of these results—including discussion of potential causes for relatively large C2C or RMSE values compared to the scanner’s nominal accuracy (e.g., registration errors, residual noise, or minor design–as-built discrepancies)—are presented in Section 4.2.

4. Research Results and Analysis

4.1. Edge Detection Results

Edge detection quality analysis is a criterion that reflects how accurately an algorithm can identify structural boundaries. In point cloud data, it is essential to correctly identify core architectural elements such as pillars, walls, and floors. Low edge detection accuracy can lead to distortions in the representation of structural components.

In this study, the performance of the RANSAC and DBSCAN algorithms was quantitatively evaluated based on the number of detected edges and the average linearity. Average linearity was calculated using the eigenvalue ratios derived from principal component analysis (PCA) for each detected edge cluster, which reflects the straightness and geometric consistency of the detected features.

These metrics enable consistent and reproducible evaluation without the need for manually defined reference data. In particular, they provide an objective basis for comparing algorithms under identical experimental conditions, ensuring that differences in performance can be clearly attributed to the underlying detection methods.

Table 2. Edge detection results.

	RANSAC		DBSCAN
	Number of Detected Edges	Average Linearity	Number of Detected Edges	Average Linearity
A-1	6	0.84	5	0.79
A-2	8	0.82	5	0.73
B-1	7	0.75	3	0.79
B-2	11	0.81	5	0.74
C-1	6	0.72	3	0.75
C-2	6	0.74	4	0.80
H-1	6	0.85	5	0.70
H-2	8	0.77	3	0.67
Average	7.25	0.79	4.13	0.75

presents the comparative results of edge detection performance between the RANSAC and DBSCAN algorithms.

As illustrated in Figure 4, RANSAC generally detected more edges and achieved equal or better linearity than DBSCAN. In particular, pillar B-2 showed the highest performance for RANSAC, with 11 detected edges and a linearity of 0.81.

RANSAC detected nearly twice as many edges on average compared to DBSCAN, indicating broader coverage of structural boundaries.

Both algorithms showed similarly high linearity scores (0.70–0.85), indicating consistent geometric precision. While DBSCAN maintained linearity with fewer detections, RANSAC achieved higher quantity without quality loss, making it more suitable when both completeness and accuracy are required.

4.2. BIM Model Comparison

To quantitatively validate the spatial accuracy of edge detection results, this section presents an evaluation based on geometric comparison with a reference BIM model. Edge point clouds extracted via RANSAC and DBSCAN were aligned to the BIM model using a two-step registration process: feature-based initial matching followed by refinement through the Iterative Closest Point (ICP) algorithm. This hybrid registration method ensured high spatial precision, with the final alignment quality confirmed by low registration RMS errors obtained after ICP refinement.

To ensure consistent interpretation, the notion of a “correct edge” was harmonized across metrics: edge points detected by each algorithm were compared with the BIM reference under a positional tolerance of τ = 10 mm. Points within this threshold were regarded as correctly detected, while those beyond were treated as deviations. All accuracy evaluations against the BIM model were conducted under a positional tolerance of τ = 10 mm. Distance-based metrics (C2C, STD, RMSE) are reported in meters, and accuracy values (where applicable) are expressed as percentages.

The accuracy assessment focuses on three indicators: mean Cloud-to-Cloud (C2C) distance, its standard deviation (STD), and root mean square error (RMSE), with all values reported in meters. The C2C distance measures the average geometric deviation between each detected edge point and the BIM surface, the STD reflects the consistency of spatial alignment across regions, and the RMSE provides a combined measure of both systematic and random deviations. These metrics allow for a continuous and detailed evaluation of geometric fidelity.

Figure 5 visualizes the spatial alignment between the detected edge points and the BIM model for the A-1 pillar. In the figure, the BIM model is displayed in gray, while the detected edge points are shown in red (RANSAC) and blue (DBSCAN). For clarity, the point sizes are slightly exaggerated to facilitate visual comparison between the two methods and the reference geometry.

This visual comparison aids in understanding how closely each algorithm’s output conforms to the true geometric boundaries of the pillar. Such visualization not only complements the numerical C2C evaluation but also highlights characteristic differences in the detection behavior of RANSAC and DBSCAN.

RANSAC exhibited a more complete and continuous distribution along the vertical boundaries of the A-1 pillar, closely following the BIM geometry. The detected edges formed dense, well-aligned contours that matched the reference model. In contrast, DBSCAN produced sparser but cleaner results, focusing on distinct vertical segments while avoiding scattered noise. However, its edge points did not fully cover the pillar’s height, indicating a more localized detection behavior. These visual differences are consistent with the numerical accuracy results in

Table 3. C2C-based accuracy evaluation between detected edges and BIM model.

Algorithm	Pillar	Mean C2C (m)	STD (m)	RMSE (m)
RANSAC	A-1	0.127	0.140	0.191
	A-2	0.099	0.123	0.168
	B-1	0.220	0.228	0.349
	B-2	0.131	0.131	0.206
	C-1	0.197	0.188	0.382
	C-2	0.251	0.185	0.398
	H-1	0.263	0.173	0.414
	H-2	0.106	0.111	0.163
DBSCAN	A-1	0.097	0.155	0.195
	A-2	0.072	0.132	0.156
	B-1	0.251	0.099	0.377
	B-2	0.129	0.094	0.171
	C-1	0.214	0.306	0.375
	C-2	0.242	0.287	0.419
	H-1	0.101	0.173	0.276
	H-2	0.102	0.116	0.171

and highlight the trade-off between RANSAC’s broader coverage and DBSCAN’s precision-focused segmentation.

Table 3 presents the mean Cloud-to-Cloud (C2C) distance, standard deviation (STD), and root mean square error (RMSE) for each pillar and algorithm. The higher deviation values compared to the nominal accuracy of the BLK360 scanner are attributed to multiple factors, including small residual misalignments after registration, differences between the as-built geometry and the idealized BIM model, the influence of algorithm parameter settings on edge fitting, and local effects of occlusions or incomplete visibility. Lower values in these indicators denote more accurate and stable edge detection. While RANSAC showed superior accuracy in certain cases (e.g., A-2), DBSCAN outperformed in others (e.g., C-1), suggesting that detection performance varies depending on the geometric characteristics of each pillar and the algorithm’s behavior. While Table 3 reports only distance-based metrics (C2C, STD, RMSE), the density-based analysis in Section 4.3 additionally incorporates Accuracy (%), calculated under the same tolerance (τ = 10 mm). These findings align with the spatial distribution patterns shown in Figure 4, where RANSAC tends to generate more continuous edge coverage, while DBSCAN exhibits cleaner yet occasionally incomplete results.

4.3. Robustness to Noise

Measurement errors, variations in reflectivity, and the presence of surrounding objects can lead to information loss when applying edge detection algorithms. Therefore, most practical experiments include a noise removal process, and this study also performed noise elimination using the Statistical Outlier Removal (SOR) algorithm. The SOR parameters were intentionally tuned to remove only 1–3% of the points, as excessive filtering could erase meaningful edge information. This reflects realistic field conditions, where aggressive noise filtering may compromise data fidelity. However, there are inherent limitations in relying solely on preprocessing algorithms to completely eliminate noise, making it essential to minimize the impact of residual noise during edge detection. Accordingly, this study included robustness to noise as a key evaluation criterion in the comparison of edge detection algorithms.

Robustness to noise refers to the extent to which an algorithm is affected by residual noise. To evaluate robustness in a manner that reflects real-world degradation, percentage-based sampling was applied. This method allows edge degradation to occur in proportion to the original distribution, enabling a more representative analysis of how each algorithm responds to noise-related sparsity without artificially flattening or uniformly distributing point density. Unlike the density analysis in Section 4.4, which applied controlled thinning percentages to simulate uniform sparsity, the noise robustness test in this section used real noise removal via SOR. These two experiments target different types of degradation and are therefore complementary. To compare this robustness, two datasets were prepared: one before and one after preprocessing with the SOR algorithm. Both the RANSAC and DBSCAN algorithms were applied to these datasets to evaluate their resistance to noise. Two performance metrics were used for the comparison: the number of detected edges and the average linearity. The number of detected edges helps assess the influence of noise on the quantity of edge detection, while average linearity serves as an indicator of how noise affects the geometric precision of detected edges.

Table 4. Number of detected edges before and after noise removal.

Algorithm	Pillar	Detected Edges (Before)	Detected Edges (After)	Δ Detected Edges
RANSAC	A-1	6	6	0
	A-2	7	8	1
	B-1	5	7	2
	B-2	9	11	2
	C-1	6	6	0
	C-2	6	6	0
	H-1	6	6	0
	H-2	8	8	0
DBSCAN	A-1	3	5	2
	A-2	3	5	2
	B-1	1	3	2
	B-2	4	5	1
	C-1	1	3	2
	C-2	2	4	2
	H-1	3	5	2
	H-2	2	3	1

and

Table 5. Average linearity before and after noise removal.

Algorithm	Pillar	Average Linearity (Before)	Average Linearity (After)	Δ Average Linearity
RANSAC	A-1	0.44	0.84	0.40
	A-2	0.45	0.82	0.37
	B-1	0.60	0.75	0.15
	B-2	0.44	0.81	0.37
	C-1	0.45	0.72	0.27
	C-2	0.47	0.74	0.27
	H-1	0.63	0.85	0.22
	H-2	0.44	0.77	0.33
DBSCAN	A-1	0.79	0.79	0.00
	A-2	0.73	0.73	0.00
	B-1	0.70	0.79	0.09
	B-2	0.74	0.74	0.00
	C-1	0.70	0.75	0.02
	C-2	0.70	0.80	0.10
	H-1	0.69	0.70	0.01
	H-2	0.64	0.67	0.03

present the changes in detected edge count and average linearity before and after preprocessing, which are visualized in Figure 6 to support comparative analysis.

To further quantify the observed changes in average linearity, a paired t-test was conducted as summarized in

Table 6. Paired t-test results on average linearity (n = 8).

Algorithm	$\mathbf{Mean} ∆$ Linearity	$95\% \mathbf{CI} \mathbf{for} ∆$	t-Statistic	p-Value
RANSAC	+0.298	[0.226, 0.369]	9.82	<0.001
DBSCAN	+0.035	[0.001, 0.069]	2.41	≈0.045

.

For RANSAC, the number of detected edges remained mostly unchanged before and after preprocessing, indicating stability in detection quantity. However, average linearity significantly improved after noise removal (up to 0.85), suggesting that RANSAC is sensitive to noise in terms of geometric precision. These results indicate that while RANSAC is not inherently robust to noise, it can produce accurate edge models when applied to clean data.

DBSCAN showed a consistent increase in the number of detected edges after noise removal across all pillars, while average linearity remained largely stable. This indicates that DBSCAN is relatively robust to residual noise, likely due to its density-based clustering mechanism, which preserves structural consistency even without preprocessing.

In summary, DBSCAN exhibited stable performance regardless of preprocessing, while RANSAC showed notable improvement in geometric accuracy after noise removal. Thus, selecting an appropriate algorithm depends on the noise characteristics of the dataset and the level of preprocessing applied.

4.4. Density-Based Performance Analysis

Point cloud data often exhibit variations in point density depending on the scanner’s position and the geometric characteristics of the target structure, and such density differences can directly affect the performance of edge detection algorithms. Typically, the quality of 3D scan data varies according to the height of the pillar, due to the distance between the scanner and the object. When the scanner is positioned on the floor, the upper part of the pillar—being farther away—tends to have lower point density, whereas the lower part—being closer—tends to have higher point density. Even under such density variations, the ability to extract edges with consistent linearity is a critical metric and plays an essential role in evaluating the reliability of an edge detection algorithm.

It should be clarified that noise robustness and density-based analyses evaluate different degradation mechanisms. The noise robustness test isolates the effect of residual measurement errors by comparing datasets before and after SOR filtering (quality degradation), whereas the density-based test investigates natural variations in point distribution caused by scanning geometry (distribution imbalance). Although both conditions alter point availability, they are conceptually distinct and thus complementary.

In addition to linearity and MDE, edge detection accuracy (%) is included in the density-based performance analysis to capture how point density variations influence the geometric correctness of detected edges. While consistency in the number of detected edges reflects algorithmic stability, it does not guarantee that these edges align accurately with the design reference. Accuracy here is calculated as described in Section 3.4.2, i.e., the proportion of detected edge points whose C2C distance to the BIM reference edge is within the positional tolerance (τ = 10 mm) relative to the total detected edge points. By including this indicator, the analysis can reveal whether density variations cause not only changes in detection quantity but also in the spatial fidelity of detected edges.

In this study, To assess algorithm performance under varying density, each pillar was divided into high- and low-density regions, defined as the top and bottom 20% of points by density. The 20% threshold was adopted to ensure clear separation between high- and low-density regions while maintaining sufficient sample size in each group for robust statistical analysis. Percentile-based grouping in this range is a common practice in point cloud and LiDAR density studies, as it provides a practical balance between contrast and representativeness [32]. Five performance metrics were analyzed: density ratio, detected edges, average linearity, mean distance error (MDE), and edge detection accuracy. Density was calculated using the inverse average distance to the 12 nearest neighbors, and stable regions were selected based on density variation.

Performance results in both density conditions for RANSAC and DBSCAN are summarized in

Table 7. Density-based performance analysis: RANSAC.

Density Region	Pillar	Detected Edges	Average Linearity	MDE (m)	Edge Detection Accuracy (%)
Low Density	A-1	8	0.95	0.354	83.7
	A-2	8	0.95	0.384	81.0
	B-1	8	0.95	0.404	81.3
	B-2	8	0.95	0.388	80.5
	C-1	8	0.95	0.271	83.6
	C-2	8	0.95	0.374	83.0
	H-1	8	0.95	0.392	82.3
	H-2	8	0.95	0.379	81.2
High Density	A-1	4	0.95	0.375	75.7
	A-2	2	0.95	0.406	64.7
	B-1	4	0.95	0.177	75.1
	B-2	8	0.95	0.356	78.2
	C-1	5	0.95	0.139	75.9
	C-2	8	0.95	0.279	78.5
	H-1	8	0.95	0.280	77.6
	H-2	8	0.95	0.219	78.3

and

Table 8. Density-based performance analysis: DBSCAN.

Density Region	Pillar	Detected Edges	Average Linearity	MDE (m)	Edge Detection Accuracy (%)
Low Density	A-1	15	0.78	0.370	84.4
	A-2	22	0.73	0.439	79.5
	B-1	21	0.75	0.414	80.3
	B-2	24	0.76	0.417	82.7
	C-1	10	0.72	0.306	82.1
	C-2	18	0.72	0.378	81.0
	H-1	28	0.71	0.385	80.6
	H-2	22	0.71	0.411	79.7
High Density	A-1	48	0.83	0.342	73.4
	A-2	59	0.78	0.342	72.2
	B-1	50	0.78	0.300	72.9
	B-2	55	0.80	0.388	77.8
	C-1	43	0.74	0.245	75.5
	C-2	48	0.79	0.321	75.0
	H-1	50	0.76	0.326	75.6
	H-2	42	0.78	0.375	77.8

.

As shown in Figure 7, RANSAC consistently detected 4 to 8 edges regardless of density, with minimal variation. This stability reflects its structure-based modeling, which is less affected by point density fluctuations.

DBSCAN detected many more edges in high-density regions—up to 59 in some cases—indicating sensitivity to dense data and a tendency for over-detection. Under low-density conditions, the number of detected edges was consistently lower across all pillars.

RANSAC maintained a fixed number of detections (often 8) in low-density regions, showing its adherence to structural modeling criteria. However, a consistent detection count does not necessarily imply high accuracy, so accuracy-based evaluation is also important.

As shown in Figure 8, RANSAC achieved higher accuracy in low-density regions, where cleaner data allowed better model fitting. In high-density conditions, accuracy dropped—especially in pillar A-2—due to false positives caused by cluttered points.

DBSCAN also showed better accuracy in low-density regions (~80%), while in high-density regions, over-detection led to a slight drop in accuracy (generally 72–78%). This reflects DBSCAN’s sensitivity to point clustering, which can result in false positives in dense areas.

Overall, both algorithms performed better in low-density regions. RANSAC maintained consistent detection counts and structural clarity, while DBSCAN, despite occasional over-detection in dense areas, preserved relatively high accuracy due to its fine-grained clustering mechanism.

Nonetheless, DBSCAN maintained relatively high accuracy even under conditions where over-detection occurred. This outcome is likely due to the nature of its density-based clustering approach, which tends to subdivide a single structural edge into multiple clusters. In essence, DBSCAN detects multiple edge segments from what would typically be interpreted as a single continuous edge in high-density regions, while RANSAC reliably extracts a stable number of structural edges, ensuring consistency in detection count.

4.5. Comparison Summary of RANSAC and DBSCAN

To synthesize the preceding analysis, this section presents a comparative summary of the RANSAC and DBSCAN algorithms based on four evaluation criteria: (1) edge detection quality, (2) BIM-based spatial accuracy, (3) robustness to noise, and (4) density-based performance. This multi-faceted approach enables a clear understanding of each algorithm’s strengths and limitations and provides practical guidance for selecting edge detection techniques in architectural point cloud processing.

The RANSAC algorithm consistently demonstrated strong structural consistency and high geometric fidelity. It maintained a stable number of edge detections across varying conditions, and its linearity significantly improved after denoising. In particular, its BIM alignment results showed more continuous and complete vertical edge distributions, relatively lower C2C and RMSE values in several pillar cases. The final alignment RMS after ICP registration also confirmed high registration quality. RANSAC’s structure-based modeling makes it highly robust to variations in both noise and point density, making it well-suited for applications requiring precise, reproducible edge detection, especially in structured environments like building columns or walls. A comparative summary of these findings is presented in

Table 9. Comparison of RANSAC and DBSCAN.

Evaluation Criteria	RANSAC	DBSCAN
Edge detection Quality Analysis	High accuracy (83.6%) with 7.25 edges on average. Suitable for structured elements.	Lower accuracy (63.9%) with 4.13 edges on average. Prone to under- or over-detection.
BIM Model Comparison	Better continuity and alignment in most pillars. Lower C2C and RMSE in some cases.	leaner alignment in some pillars. Occasionally better mean C2C, but RMSE tends to be higher.
Robustness to noise	Linearity improved (+0.30), edge count maintained. Statistically significant improvement after denoising confirms RANSAC’s post-processing advantage.	Minimal change in linearity (+0.04), but consistent increase in edge count across all pillars. Statistically minor yet stable performance under noise.
Density-based Performance Analysis	Accuracy improved (75.7% → 83.6%), consistent edge count. Robust to density variation.	Edge count increased (15 → 59), but accuracy decreased (84.4% → 72.2%). Sensitive to density.

.

In contrast, DBSCAN responded more sensitively to changes in environmental conditions. While its average linearity was slightly lower and statistically less pronounced, it excelled in detecting clean and well-defined segments, particularly under low-density conditions, where over-detection was suppressed. Although DBSCAN occasionally outperformed RANSAC in terms of mean C2C distances, especially in cases like C-1 and B-1, its RMSE values were often higher, reflecting a less consistent spatial fit. It also struggled to maintain detection accuracy under high-density conditions due to its tendency to split continuous edges into multiple clusters. Nevertheless, its performance was remarkably stable even without preprocessing, reflecting robustness to residual noise. These trends were statistically validated through paired t-tests on linearity scores, confirming RANSAC’s significant improvement after denoising, while DBSCAN exhibited small but consistent changes.

In conclusion, RANSAC is ideal for tasks requiring structural restoration and consistent detection, while DBSCAN is more adaptable for detecting subtle geometric variations and noise-prone environments. These algorithms can be applied complementarily, and algorithm selection should consider data characteristics and project goals for optimal performance.

5. Discussion

This study provides an empirical basis for algorithm selection by comparing RANSAC and DBSCAN under consistent experimental conditions using identical structural data. By quantitatively analyzing edge detection characteristics—such as the number of detected edges, linearity, BIM-based geometric accuracy, and mean distance error (MDE)—the study goes beyond simply ranking algorithm performance and instead explores their suitability and limitations under varying architectural conditions.

Taken together, the findings indicate that RANSAC is well-suited for edge detection tasks requiring accuracy, continuity, and reproducibility. Its stable detection count and high linearity—even under density variations—align with the expectations of geometry-based methods applied in prior studies. For instance, He et al. [15] and Wang et al. [14] also reported strong fitting performance of RANSAC when used in structured environments such as roads or agricultural boundaries. This study extends those findings to interior architectural components, confirming RANSAC’s robustness under real-world conditions. In particular, a paired t-test revealed a statistically significant improvement in average linearity after denoising (Table 6), further validating its sensitivity to noise and benefit from preprocessing.

On the other hand, DBSCAN showed higher sensitivity to environmental changes, particularly point density, but retained stable performance in noisy datasets. This supports previous findings by Chen et al. [16] and Guo et al. [17], who noted DBSCAN’s effectiveness in clustering high-density features amidst noise, albeit with a tendency toward over-segmentation. Although its improvement in linearity was statistically significant, the change was marginal, suggesting that DBSCAN is inherently robust to residual noise even without preprocessing. Our results similarly show that DBSCAN performs well in low-density regions but tends to over-detect in high-density areas—thereby decreasing overall accuracy.

While RANSAC’s stability under noise is consistent with previous studies [14,15], its sensitivity to repetitive rectilinear geometries in indoor environments was more pronounced than reported in outdoor or infrastructural contexts. Similarly, although DBSCAN’s robustness to noise supports prior findings [16,17], our experiments revealed a stronger tendency toward over-detection in high-density regions, which was less emphasized in earlier works. These discrepancies suggest that algorithm behavior is not only method-dependent but also strongly influenced by architectural settings and scanning conditions.

Importantly, this study contributes a new layer of analysis by evaluating spatial alignment against a CAD-derived BIM reference proxy (Section 4.2), which has been insufficiently addressed in earlier works. Unlike studies that rely solely on geometric indicators such as curvature or normal vectors [9,11], the inclusion of BIM-derived C2C and RMSE evaluation strengthens the practical applicability of the findings for design-to-scan comparison and BIM integration tasks. While the model serves as a consistent baseline for relative evaluation, it is acknowledged that it may deviate from the as-built geometry. To confirm its reliability, several representative points were spot-checked during scanning using a laser distance meter, all of which were within the device’s tolerance (±2 mm), supporting its suitability as a comparative reference. Additionally, the final registration RMS error obtained after ICP refinement confirmed that spatial alignment between scan data and BIM was within acceptable limits.

Nevertheless, absolute deviations—such as C2C distances exceeding the scanner’s nominal precision—were observed, reflecting the inherent differences between idealized BIM geometry and real-world conditions. These were accounted for during analysis by focusing on relative algorithm performance across identical conditions.

In summary, rather than determining the superiority of one algorithm, the study emphasizes complementary strengths. RANSAC excels in structural regularity and consistent model fitting, while DBSCAN offers flexibility in detecting subtle or irregular features. These insights echo Scitovski et al. [28], who proposed hybridizing RANSAC and DBSCAN for multi-shape detection. Although our study does not combine the two algorithms, it validates their complementary potential through systematic benchmarking.

By incorporating statistical evidence, spatial accuracy evaluation, and sensitivity to noise and density, this study provides a reproducible framework for assessing 3D edge detection in architectural contexts. It sets a foundation for follow-up research involving more complex objects, learning-based techniques, and integration with digital construction workflows.

6. Conclusions and Limitations

This study conducted a comparative evaluation of RANSAC and DBSCAN for edge detection using 3D point cloud data acquired from eight rectangular interior columns. All scans were collected with a terrestrial laser scanner, and the point clouds were preprocessed using the Statistical Outlier Removal (SOR) algorithm. The two algorithms were evaluated using four performance criteria: edge detection quality, BIM-based spatial accuracy, robustness to noise, and density-based performance.

The experimental results revealed distinct characteristics for each algorithm. RANSAC consistently detected a stable number of structural edges with high linearity, even under varying noise and density conditions. Its average linearity improved significantly after noise removal, as confirmed by statistical testing (Table 6), and its detected edges exhibited better alignment with the BIM reference in most cases. These findings demonstrate RANSAC’s suitability for structured environments where reliable geometric fitting and spatial consistency are essential.

DBSCAN, in contrast, showed a higher degree of sensitivity to point density and environmental variability. In high-density regions, it tended to over-segment features, reducing overall accuracy, while in low-density regions, it demonstrated more controlled detection behavior. Although its linearity improvement was statistically valid, the magnitude of change was relatively small, suggesting inherent resilience to residual noise even without preprocessing. This suggests that DBSCAN may be better suited for applications that prioritize sensitivity or require flexibility in less controlled scanning conditions.

The incorporation of BIM-based geometric evaluation metrics, including Cloud-to-Cloud (C2C) distance and RMSE, added a valuable geometric accuracy perspective. Notably, DBSCAN occasionally produced lower C2C values in certain pillars, indicating that although it detected fewer features, some were geometrically closer to the reference model. However, it is important to note that the absolute deviation values were larger than the scanner’s nominal accuracy, reflecting the gap between idealized BIM geometry and real-world conditions. This trade-off between structural coverage and local fit precision reinforces the importance of selecting algorithms based on project-specific accuracy needs and data conditions. Nevertheless, these geometric metrics primarily capture spatial deviation and alignment, while offering limited insight into edge continuity or semantic correctness, which constrains their interpretability for broader applications.

However, this study has several limitations. All test data were limited to rectangular interior columns, which, while offering controlled experimental conditions, do not fully represent the diversity of architectural geometries. Furthermore, although the BIM model provided a design-based reference proxy for alignment evaluation, it was developed from design drawings rather than precise as-built surveying. During scanning, several representative points were spot-checked using a laser distance meter to confirm dimensional consistency with the actual construction, and all were within the device’s tolerance (±2 mm). Nonetheless, small discrepancies between the BIM model and the as-built geometry may still exist. In addition, the reliance on Statistical Outlier Removal (SOR) preprocessing and manual spot-checking introduces procedural subjectivity, which may affect the reproducibility and generalizability of the results. Moreover, the current evaluation focused primarily on geometric metrics, without incorporating semantic or functional analysis of detected features. While this limits interpretability for certain applications, the geometric evaluation—especially BIM-based C2C comparison—offers a practical baseline for construction quality control.

These limitations underscore that the present findings should be interpreted as preliminary evidence rather than universal conclusions, and they directly motivate the following directions for future research. For future research, we plan to extend the evaluation framework to cover a wider range of structural types, including curved walls, complex joints, and mechanical systems. Additionally, the incorporation of learning-based edge detection methods will be explored to improve adaptability in dense or occluded environments. Finally, by integrating real-time edge detection with automated BIM-based quality control systems, we aim to contribute to scalable and precise digital construction workflows.