SPICD-Net: A Siamese PointNet Framework for Autonomous Indoor Change Detection in 3D LiDAR Point Clouds

Abstract

Reliable change detection in indoor environments remains a challenge for autonomous robotic systems using 3D LiDAR. Existing methods often require manual annotation, computationally intensive architectures, or focus on outdoor scenes. This paper presents SPICD-Net, a lightweight Siamese PointNet framework for indoor 3D change detection trained exclusively on synthetically generated anomalies, eliminating manual labeling. The framework offers three deployment-oriented contributions: a three-class Siamese formulation separating no-change, changed, and geometrically inconsistent tile pairs; a pre-FPS anomaly injection strategy that aligns synthetic training with inference-time preprocessing; and a stochastic-gated Chamfer-statistics branch that complements learned embeddings with explicit geometric cues under consumer-grade hardware constraints. Evaluated on 14 controlled simulation experiments in an indoor corridor dataset, SPICD-Net achieved aggregated Precision = 0.86, Recall = 0.82, F1-score = 0.84, and Accuracy = 0.96, with zero false positives in the no-change baseline and mean inference time of 22.4 s for a 172-tile map on a single consumer GPU. Additional robustness experiments identified registration accuracy as the main operational prerequisite. A limited real-world validation in one unseen room (four scans, 67 tiles) achieved Precision = 0.583, Recall = 1.000, and F1 = 0.737.

1. Introduction

Autonomous mobile robots deployed for inspection, logistics, and facility management depend on accurate, up-to-date representations of their operating environment. In indoor settings such as warehouses, factory floors, and hospital corridors, the physical layout can change frequently: furniture is rearranged, equipment is installed or removed, and temporary obstructions appear and disappear. When a robot’s internal map ceases to reflect the true state of the environment, localization degrades, path planning fails, and safety constraints may be violated. Detecting and localizing such changes automatically is therefore a prerequisite for long-term autonomous operation in dynamic indoor spaces.

Three-dimensional point clouds acquired by LiDAR sensors offer a rich geometric representation for change detection, providing direct depth measurements that are independent of lighting conditions and texture ambiguity [1,2]. Unlike 2D image-based approaches, which must infer 3D structure from perspective projections and are sensitive to viewpoint, illumination, and appearance variation, point cloud methods operate directly on the spatial geometry of the scene. The proliferation of affordable LiDAR units on mobile robot platforms, including the Unitree L1 sensor used in this study, has made dense 3D scanning increasingly accessible for routine environment monitoring.

Despite recent advances in 3D change detection, a review of the literature (Section 2) reveals three interrelated gaps that none of the existing methods fully addresses. First, deep Siamese architectures for 3D change detection have been developed and evaluated exclusively on outdoor urban environments captured by aerial or vehicular LiDAR [1,3,4,5]; indoor mobile LiDAR change detection remains largely unexplored. Second, all state-of-the-art supervised methods rely on manually annotated point cloud pairs [3,4,5], an annotation effort that is impractical for autonomous robotic deployment (e.g., the HKCD dataset required three months and three annotators [5]); although synthetic pre-training reduces this burden [3], a fully annotation-free indoor pipeline has not been proposed. Third, current high-performance architectures require server-class GPUs (Tesla V100) [3,5,6], making them unsuitable for on-robot or workstation-based inference where compute is constrained.

To address these gaps, this paper presents SPICD-Net and makes the following contributions:

• An end-to-end indoor mobile LiDAR change detection pipeline that reformulates the problem around raw scan registration, fused map construction, volumetric tiling, fixed-size patch extraction, Siamese inference, and spatial projection of detected changes for indoor robotic environments rather than outdoor urban scenes;
• A pre-FPS anomaly injection strategy that generates realistic training pairs entirely from synthetic changes (point addition, removal, or displacement) applied to dense reference tiles before Farthest Point Sampling, thereby eliminating manual annotation while reducing the mismatch between synthetic training pairs and inference-time preprocessing;
• A deployment-ready system integration combining an existing Siamese PointNet backbone with an auxiliary Chamfer-statistics branch and stochastic gating, representing a purposeful combination of established components optimized for indoor tile-level change detection on consumer-grade hardware (GTX 1650, 4 GB VRAM) rather than a novel backbone architecture, and achieving an F1-score of 0.84 with zero false positives in the no-change baseline condition.

The remainder of this paper is organized as follows. Section 2 reviews related work on 3D point cloud change detection, point cloud deep learning, and Siamese architectures. Section 3 details the proposed methodology. Section 4 describes the experimental setup. Section 5 presents the results. Section 6 discusses the findings and limitations. Section 7 concludes the paper.

2. Related Work

This section reviews the five areas of prior work most relevant to the proposed framework: classical and deep 3D change detection, deep learning on point sets, Siamese architectures for 3D tasks, point cloud registration, and geometric distance metrics with data augmentation.

2.1. 3D Change Detection in Point Clouds

Change detection in 3D point clouds has been studied extensively in the photogrammetry and remote sensing communities, as surveyed by Stilla and Xu [1] and Kharroubi et al. [2]. Classical approaches typically compute geometric differences between bi-temporal scans and classify changes by thresholding a distance-based score. Girardeau-Montaut et al. [7] introduced cloud-to-cloud (C2C) distance and occupancy-grid comparison for terrestrial laser scanning data, establishing foundational geometric baselines that remain widely used in practice through tools such as CloudCompare (v2.13.2). More refined methods, including multiscale model-to-model cloud comparison (M3C2), estimate distances along local surface normals and provide per-point confidence intervals, improving robustness to varying point density and residual registration error [1].

These geometric methods are attractive because they require no training data and are easy to deploy once scans are registered. For this reason, they remain important reference baselines in practical 3D change detection workflows. However, they also exhibit several limitations: sensitivity to registration residuals, inability to distinguish genuine changes from noise, and lack of semantic understanding of the detected differences [1,2].

The vast majority of 3D change detection research targets outdoor urban environments, building facades, vegetation, terrain, and street furniture, acquired by aerial or mobile LiDAR [1,3,4,5].

Indoor change detection has received comparatively little attention in the point cloud literature. Qiu et al. [8] recently proposed a joint change detection and captioning framework for indoor scenes using RGB-D data, but their approach relies on pre-segmented object instances and camera-based reconstruction rather than direct LiDAR point cloud comparison.

In the present study, C2C and M3C2 are therefore used not only as representative prior methods from classical 3D change detection, but also as training-free geometric baselines in our benchmark evaluation under a unified indoor tile-level protocol.

2.2. Deep Learning on Point Sets

The emergence of deep learning architectures for unstructured point sets has fundamentally changed 3D scene understanding. PointNet [9] demonstrated that a shared multi-layer perceptron applied independently to each point, followed by a symmetric aggregation function (global max pooling), can achieve permutation invariance while learning effective features for classification and segmentation. Owing to its simplicity, efficiency, and low computational cost, PointNet remains particularly relevant for lightweight indoor robotics applications and therefore also serves as an important benchmark family for our study.

However, PointNet’s reliance on global max pooling means that it does not capture local geometric structures induced by the metric space in which points reside [10]. PointNet++ [10] addressed this limitation by introducing hierarchical set abstraction layers that recursively apply PointNet to local neighborhoods defined by ball queries around farthest-point-sampled centroids. This hierarchical design captures fine-grained local patterns at multiple scales and significantly improves performance on tasks requiring local sensitivity, such as semantic segmentation and scene understanding. Multi-scale and multi-resolution grouping strategies further enable the network to adapt to non-uniform point densities. In the present study, a Siamese PointNet++ variant is also included in the benchmark comparison as a stronger hierarchical learned baseline against which the proposed lightweight PointNet-based design can be assessed.

Beyond clean-data performance, robustness under realistic perturbations is also critical for point-cloud learning. Ren et al. [11] systematically benchmarked point-cloud classification architectures under various corruptions, including noise, density variation, and geometric perturbations, demonstrating that while modern architectures achieve high clean accuracy, their performance degrades substantially under realistic conditions. This observation is directly relevant to the present work, where a lightweight PointNet-based Siamese design is adopted for efficient indoor deployment and compared against alternative geometric and learned baselines under a unified evaluation protocol. This finding motivates the use of targeted augmentation strategies during training, as adopted in our framework (Section 3.8).

2.3. Siamese Networks for 3D Change Detection

Siamese architectures, which process two inputs through shared-weight encoders and compare the resulting feature representations, are a natural fit for change detection. De Gélis et al. [3] introduced Siamese KPConv, the first deep Siamese network for multi-class change segmentation directly on raw 3D point clouds. The architecture employs Kernel Point Convolution (KPConv) encoders with shared weights, a nearest-point feature difference fusion strategy, and a decoder that produces pointwise change labels. Evaluated on both synthetic (Urb3DCD-V2) and real (AHN-CD) urban datasets, Siamese KPConv achieved approximately 80% Mean Intersection over Union (mIoU) over change classes, outperforming traditional methods by 15–30% in mIoU. Notably, the authors demonstrated that pre-training on simulated data can reduce real-data annotation requirements by over 3000×, a finding that motivates our fully synthetic training approach. Because KPConv-based Siamese architectures represent a strong but computationally heavier family of learned 3D change detection models, we also include a Siamese-KPConv benchmark variant in our comparative evaluation to contrast raw detection performance with deployment cost under the same indoor tile-level protocol.

In a follow-up study, de Gélis et al. [4] proposed three evolved architectures: OneConvFusion, Triplet KPConv, and Encoder Fusion SiamKPConv, that incorporate change-related features at different stages of the encoder–decoder pipeline. The Encoder Fusion variant achieved the highest performance by fusing bi-temporal features early in the encoding process. The study also introduced hand-crafted features, notably the Stability feature, which significantly improved the detection of specific change classes.

Han et al. [6] proposed PBFormer, a transformer-based Siamese architecture that uses point transformer encoders and a bi-spatiotemporal fusion module for pointwise change detection. Xiao et al. [5] introduced PGN3DCD, which incorporates non-parametric 3D change prior knowledge (algebraically computed change masks) to guide the neural network toward change-related features, achieving state-of-the-art results on both the Urb3DCD-V2 and the newly constructed HKCD real-world dataset. The SHREC 2023 benchmark [12] evaluated multiple Siamese approaches for object-level change classification in city scenes, confirming that Siamese architectures with geometric feature exploitation are effective for this task and that synthetic training data can narrow the domain gap with real data.

All of these methods target outdoor urban environments at pointwise segmentation granularity and require annotated training data. Our work differs in three respects: we target indoor environments, operate at tile-level classification granularity, and train entirely on synthetically generated anomalies without any manual annotation.

Beyond change detection, Siamese frameworks have also proven effective for 3D single object tracking [13,14,15,16], where a template point cloud is compared against a search region across frames—a formulation directly analogous to the reference-tile versus current-tile comparison in the present work. This confirms that Siamese networks can learn discriminative representations for comparing 3D point cloud regions, reinforcing their suitability as the architectural foundation of SPICD-Net.

2.4. Point Cloud Registration

Accurate alignment of bi-temporal scans is a prerequisite for any point cloud change detection method. The Iterative Closest Point (ICP) algorithm [17] remains the most widely used approach, alternating between correspondence estimation and least-squares rigid transformation optimization. Point-to-plane ICP variants, which minimize the distance along the surface normal, offer improved convergence for planar environments such as indoor corridors.

Learning-based registration methods have emerged as powerful alternatives. PointNetLK [18] combines PointNet feature extraction with the Lucas–Kanade alignment algorithm to perform registration without explicit correspondence estimation, achieving computational efficiency and noise robustness. PCRNet [19] extends this concept using a PointNet-based Siamese architecture to directly regress the transformation aligning two feature vectors. Deep Closest Point (DCP) [20] incorporates an attention mechanism to learn soft correspondences between point clouds, outperforming ICP on ModelNet40 benchmarks.

For large-scale scenes, feature-based methods have proven essential. 3DMatch [21] learns local geometric descriptors from RGB-D reconstructions for feature matching, while Fully Convolutional Geometric Features (FCGF) [22] and the Perfect Match [23] learn dense per-point descriptors that enable robust correspondence estimation. PREDATOR [24] specifically targets low-overlap registration by learning to attend to the overlap region, achieving 90.6% recall on the 3DMatch benchmark. Deep Global Registration (DGR) [25] integrates learned features with a six-DOF optimization layer, achieving 91.3% recall while remaining competitive in inference time.

Our framework employs classical point-to-plane ICP for scan-to-map registration, as implemented in our prior work on automated map construction from mobile LiDAR scans [26]. This choice prioritizes reliability and simplicity for the indoor corridor setting, where high initial overlap between consecutive scans and planar geometry ensures rapid ICP convergence.

2.5. Geometric Distance Metrics and Data Augmentation

The Chamfer distance is a fundamental metric for comparing unordered point sets, computing the average nearest-neighbor distance in both directions. Standard Chamfer distance treats all point pairs equally, making it sensitive to outliers. Wu et al. [27] proposed Density-Aware Chamfer Distance (DCD), which incorporates a weighting mechanism based on local point density to address the imbalanced contribution of dense and sparse regions. Montanaro et al. [28] introduced Hyperbolic Chamfer Distance (HyperCD), which computes Chamfer distance in hyperbolic space to assign higher gradient weight to well-matched pairs while gradually improving poorly matched ones. These variants inform our use of Chamfer statistics as an auxiliary feature branch (Section 3.6.3), where we extract statistical summaries (mean, max, median, 90th percentile) of bidirectional nearest-neighbor distances rather than a single scalar distance.

Data augmentation is critical for training robust point cloud models, particularly when training data are limited or synthetically generated. Lee et al. [29] proposed RSMix, a rigid mixed-sample augmentation that exchanges local regions between point clouds while preserving geometric consistency. Kim et al. [30] introduced PointWOLF, which applies spatially varying non-rigid deformations through locally weighted transformations. Xu et al. [31] demonstrated that diverse synthetic data, combined with a learned abstaining penalty, enables effective LiDAR outlier detection without real-world outlier labels.

These works collectively support our design decision to train exclusively on synthetically generated anomalies with targeted augmentation (jitter, dropout, pose perturbation), rather than requiring manually annotated real-world change examples.

2.6. Research Gaps and Contributions of This Work

The foregoing review confirms that none of the existing methods fully addresses the combination of indoor domain, annotation-free training, and resource-constrained deployment that motivates the present work. The three contributions introduced in Section 1 directly target these gaps, and the methodology implementing them is described in detail in Section 3.

3. Methodology

3.1. Problem Formulation

We formulate indoor change detection as a supervised classification problem over pairs of 3D point cloud patches. Let

$$P_{ref}={\left\{x_i\in {\mathbb{R}}^3\right\}}_{i=1}^N,$$

(1)

denote a reference point cloud patch and

$$P_{cur}={\left\{x_j\in {\mathbb{R}}^3\right\}}_{j=1}^N,$$

(2)

a current observation of the same spatial region, each containing N = 4096 points. The goal is to learn a function

$$f_{\theta }:(P_{ref}, P_{cur})\to \{0,1,2\},$$

(3)

that classifies each pair into one of the three categories: Class 0 (same tile, no change), Class 1 (same tile, change detected), or Class 2 (different tiles). This three-class design departs from the binary formulations common in existing 3D change detection literature [3,4]. By explicitly introducing a dedicated “different tile” class as a hard negative, the network is required to distinguish genuine geometric changes from the larger structural dissimilarities that arise when unrelated spatial regions are compared, thereby sharpening the decision boundary between Classes 0 and 1.

3.2. System Overview

The proposed framework comprises three stages, illustrated in Figure 1.

In Stage 1 (Reference Map Construction), mobile LiDAR scans collected by a Unitree Go2 quadruped robot are registered into a unified global coordinate frame using the automated registration pipeline presented in our prior work [26], which performs overlap detection, point-to-plane ICP [17], pose graph optimization, and voxel downsampling to 5 cm resolution. The fused map is then partitioned into a volumetric grid of spatial tiles, and each tile is subsampled to a fixed-size reference patch via Farthest Point Sampling (FPS).

In Stage 2 (Offline Training), synthetic anomalies—object additions, removals, and displacements—are injected into dense tiles before FPS to generate labeled training pairs. These pairs, together with no-change and different-tile pairs, are used to train a Siamese PointNet classifier with focal loss. In Stage 3 (Online Inference), a new LiDAR scan is registered to the reference frame and partitioned identically into tiles. Each tile is compared with its reference patch by the trained model. Tiles whose Class 1 probability exceeds the F1-optimized threshold are flagged as changed, producing a spatial heatmap of detected changes. When a physical object extends across multiple neighboring tiles, we interpret the resulting activations jointly as a single spatially extended change region rather than as isolated detections.

3.3. Data Acquisition and Preprocessing

3.3.1. LiDAR Data Collection

Point cloud data were acquired using a Unitree Go2 quadruped robot equipped with a Unitree 4D LiDAR L1 sensor, which provides 360° horizontal and 90° vertical field of view at an effective sampling rate of 21,600 points per second [26]. The robot traversed an indoor corridor environment along planned paths, collecting 42 scans that collectively cover the full extent of the monitored space (25 scans were used to create the global map). Each raw scan contains approximately 10⁵ points representing walls, floors, ceilings, doors, and objects within the corridor.

3.3.2. Point Cloud Registration

Individual scans are aligned into a common global coordinate frame through the multi-stage registration pipeline described in [26]. First, automatic overlap detection eliminates non-overlapping scan pairs by comparing bounding-box geometry, centroid Euclidean distances, and height differences, reducing the number of Iterative Closest Point (ICP) calls by over 80%. Remaining candidate pairs undergo point-to-plane ICP [17], which iteratively minimizes the point-to-plane distances:

$$\mathrm{E}(\mathbf{R},\mathit{t}) = {\sum }_{\mathit{i}}{\left[(\mathbf{R}{\mathit{x}}_{\mathit{i}}+t-{\mathit{y}}_{\mathit{i}})· {\mathit{n}}_{\mathbf{i}}\right]}^2,$$

(4)

where x_i and y_i denote corresponding point pairs, n_i is the surface normal at y_i, and $(\mathit{R},\mathit{t})\in \mathit{S}\mathit{E}\left(\mathbf{3}\right)$ represents the rigid transformation. Registrations that do not meet quality thresholds (fitness > 0.55, Root Mean Square Error (RMSE) < 0.05 m) are discarded. A pose graph is then constructed over all valid pairwise registrations, where nodes represent scans and edges represent successful transformations. The largest connected component is identified, and a reference scan is selected based on the highest node degree and mean fitness. All other scans are brought into the reference coordinate frame via multi-hop transformation chaining along the shortest graph paths [26], achieving an average alignment RMSE of approximately 2.5 cm.

To improve methodological transparency and reproducibility, the principal ICP and registration-pipeline settings used in Stage 1 are summarized in

Table 1. Registration pipeline settings used in Stage 1.

Parameter	Value
ICP variant	Point-to-Plane
Correspondence distance threshold	0.05 m
Normal estimation radius	0.10 m
Normal estimation max neighbors	30
Acceptance threshold—fitness	>0.55
Acceptance threshold—RMSE	<0.05 m
Voxel size before ICP	0.05 m
Initial transformation	Identity $(I_4$)

.

Registration quality statistics for the Stage 1 map-construction pipeline are reported in

Table 2. Registration quality statistics for the Stage 1 map-construction pipeline.

Statistic	Value
Total scans collected	42
Successful ICP registrations	25 scans in main connected component
Discarded registrations	17 scans
Average alignment RMSE (main component)	2.50 ± 0.45 cm
Average fitness (main component)	0.592 ± 0.089

.

In the present pipeline, pairwise scan alignments are estimated with point-to-plane ICP on voxel-downsampled point clouds using a 5 cm correspondence threshold and surface normals estimated within a 10 cm neighborhood. Registrations are accepted only if they satisfy the pipeline’s quality criteria; otherwise, they are excluded from the pose graph before global refinement. Of the 42 collected scans, 25 formed the main connected component after registration filtering, whereas 17 were discarded because the resulting alignments did not meet the required fitness and RMSE thresholds. The retained registrations achieved an average alignment RMSE of 2.50 ± 0.45 cm and an average fitness of 0.592 ± 0.089.

3.3.3. Map Fusion and Voxel Downsampling

Registered scans from the main graph component are merged into a single global point cloud, following the procedure in [26]. To normalize point density and reduce computation burden, the fused map is voxel downsampled at a 5 cm resolution using Open3D, illustrated in Figure 2. This resolution was chosen to match the inference preprocessing pipeline exactly, ensuring consistency between the reference representation and the current scan observations at test time.

A finer voxel size would preserve more geometric detail but increase map density, memory usage, and registration or inference cost, whereas a coarser voxel size would suppress subtle structural changes near the detection floor. The selected 5 cm resolution therefore represents a practical compromise between geometric fidelity and computational efficiency and matches the preprocessing applied at inference time.

3.4. Spatial Tiling and Patch Extraction

The global map is partitioned into a regular volumetric grid with tile dimensions $L_x\times L_y\times L_z=1.5\times 1.5\times 1.0 \mathrm{m}$. These dimensions were selected empirically to balance spatial granularity against the requirement that each tile contain sufficient geometric structure. This design entails an inherent trade-off: smaller tiles improve spatial localization of changes but increase patch sparsity and reduce geometric stability, whereas larger tiles improve robustness at the cost of coarser localization. Tile-level processing is therefore used as a practical compromise for lightweight indoor change detection on consumer-grade hardware. Tiles containing fewer than 512 points are discarded as uninformative. For each retained tile, FPS selects exactly N = 4096 points; FPS iteratively chooses the point farthest from the already selected set, yielding spatially uniform coverage that is superior to random subsampling for preserving geometric detail [9]. In tiles with fewer than 4096 points, existing points are repeated to reach the target count. The FPS output for each tile constitutes the stored reference patch; tile metadata (spatial indices, 3D center coordinates, and bounding box) are stored in a JSON manifest for downstream use. For reproducibility, the structure of the JSON manifest used for tile-to-patch mapping is summarized in Appendix A 

Table A1. JSON manifest schema for tile-to-patch mapping.

Field	Description	Units/Type
tile_size	Dimensions of each spatial tile in the X, Y, and Z directions	3-element float array, meters; here [1.5, 1.5, 1.0]
patch_size	Target number of points retained per tile after FPS	integer; here 4096
min_points	Minimum number of raw points required for a tile to be retained	integer; here 512
save_dense_tiles	Indicates whether dense pre-FPS tile files were stored	boolean
dense_voxel	Optional voxel size applied to dense stored tiles; −1 indicates none	float, meters
map_bounds.min	Minimum corner of the fused global map bounding box	3-element float array, meters
map_bounds.max	Maximum corner of the fused global map bounding box	3-element float array, meters
n_tiles_grid	Number of tiles along each grid axis	3-element integer array; here [14, 13, 2]
patches	List of retained valid patches that passed the minimum-point filter	array of patch records
patches[].patch_id	Unique patch identifier derived from zero-padded tile indices	string, e.g., tile_III_JJJ_KKK
patches[].tile_idx	Integer grid indices of the tile	3-element integer array
patches[].tile_center	Tile center in global reference coordinates	3-element float array, meters
patches[].n_original_points	Number of raw points in the tile before FPS	integer; in this map 512–6392
patches[].n_sampled_points	Number of output points after FPS/padding	integer; always 4096
patches[].file	Relative path to the stored reference patch	string; npy, shape [4096, 3], float 32
patches[].dense_file	Relative path to the stored dense pre-FPS tile	string; npy, float 32
n_valid_patches	Number of tiles retained after filtering	integer; here 172
n_skipped_tiles	Number of tiles discarded for having fewer than min_points	integer; here 192

.

3.5. Synthetic Change Generation

A central design choice of this framework is the exclusive use of synthetically generated training labels, eliminating the need for manually annotated real-world change data. This strategy is motivated by two practical considerations: (i) collecting temporally separated indoor LiDAR scans with known ground-truth changes is expensive and difficult to scale, and (ii) synthetic anomalies can be parametrically controlled, enabling systematic variation in change type, magnitude, and density during training.

This choice enables systematic, fully annotation-free training but also means that the present study evaluates controlled synthetic changes rather than real-world changes with public ground-truth labels.

3.5.1. Anomaly Operations

Three atomic operations simulate common indoor changes. Each operates on a dense tile (before FPS) centered at a randomly selected point and is parametrized by a magnitude (spatial displacement in meters) and a density (fraction of affected points).

Add object: Simulates the appearance of a new object. The $k =\left[d·N_{dense}\right]$ nearest points to the anomaly center are duplicated, shifted by a random 3D displacement of norm equal to the specified magnitude, and perturbed with Gaussian jitter $\sigma =0.001 + 0.05 · m$, where m is the magnitude. The vertical component of the shift is attenuated by a factor of 0.25 to reflect the predominantly horizontal layout of indoor objects. Added points are clipped to remain within the tile bounds.

Remove object: Simulates the disappearance of an object. The k nearest points to the anomaly center are deleted from the tile, subject to a minimum floor of 64 retained points to prevent degenerate patches.

Move object. Simulates the displacement of an existing object. The k nearest points are translated along a random direction by the specified magnitude. The vertical component is again attenuated (factor 0.3), and the moved points are clipped to the tile bounding box.

During training, the magnitude is drawn uniformly from [0.10, 0.50] m, and the density is fixed at 7.5% of the tile point count. One of the three operations is selected uniformly at random for each changed pair.

The synthetic-generation parameters were selected to approximate plausible indoor geometric changes while remaining consistent with the spatial resolution of the pipeline. In particular, the training magnitude range [0.10, 0.50] m spans displacements from near the effective detection floor of the 5 cm voxelized representation to clearly visible object-scale movements within the corridor environment. The additive Gaussian jitter was introduced to avoid unrealistically rigid duplicated structures, while keeping the perturbation magnitude proportional to the injected displacement. Vertical displacement components were attenuated to reflect the fact that most indoor scene changes in the present setting, such as moved chairs, boxes, or equipment, are dominated by horizontal repositioning rather than large vertical motion.

3.5.2. Pre-FPS Anomaly Injection

A critical implementation detail is that anomalies are injected into the dense tile representation before FPS rather than into the already subsampled 4096-point reference patch. This pre-FPS strategy ensures that the spatial distribution of anomalous points undergoes the same FPS downsampling that occurs during inference, when a newly acquired scan is sampled from its raw density. Post-FPS injection, by contrast, would introduce points at an unrealistic density relative to the surrounding geometry, creating an artificial statistical signature that the network could exploit as a shortcut rather than learning genuine geometric change cues. Adopting pre-FPS injection therefore better reflects inference-time conditions and was observed to improve robustness under the synthetic evaluation protocol. Figure 3 illustrates each operation applied to an example reference tile.

3.5.3. Training Pair Composition

Each training batch comprises three categories of pairs, sampled at fixed ratios:

• No-change pairs (50% of the batch) present the same tile observed under different conditions. To prevent the model from relying on point-for-point identity, 70% of no-change pairs are generated by re-running FPS with a different random seed on the dense tile, followed by shared nuisance transforms (random yaw rotation, small translation, scan dropout of 10–50% of points, and residual pose jitter of $\sigma =0.001 + 0.05 · m$). The remaining 30% use the stored reference patch with only negligible jitter ($\sigma =0.0001$), closely mimicking the inference setting where the reference patch is loaded directly from the disk.
• Changed pairs (35%) present a reference patch and a synthetically modified current patch, as described above, and are labeled Class 1. Such nuisance augmentations are applied at half the intensity of no-change pairs to preserve the change signal.
• Different-tile pairs (15%) pair patches from two spatially separated tiles (minimum 3 m apart) and are labeled Class 2. These hard negatives explicitly train the network to treat large geometric dissimilarity as geometric inconsistency rather than as change, which is essential for reducing false positives in spatially heterogeneous indoor environments.

The training density was fixed at 7.5% to bias learning toward relatively subtle local changes rather than only large, visually obvious perturbations. This produces a conservative training regime, while the denser simulation experiments in Section 4 (10–50%) evaluate whether the learned representation remains sensitive as the proportion of affected points increases. The allocation of 15% of each batch to different-tile pairs was chosen to provide a consistent stream of hard negatives, making the separation between true intra-tile change and unrelated spatial mismatch an explicit part of the learning problem. The minimum separation of 3 m was imposed to reduce accidental local overlap between such pairs and to ensure that Class 2 examples represent genuinely inconsistent spatial contexts.

3.6. SPICD-Net Architecture

The proposed SPICD-Net (Siamese PointNet for Indoor Change Detection) architecture, illustrated in Figure 4, implements the task-specific formulation introduced in Section 3.1 through a shared-weight Siamese design in which two point cloud patches are processed by a PointNet encoder [9], and the resulting feature vectors are fused and classified. The contribution is system-level and deployment-oriented: it lies not in introducing a new backbone, but in the purposeful integration of a lightweight Siamese PointNet encoder with explicit geometric cues, a three-class formulation, and synthetic training, specifically designed for annotation-free indoor deployment on consumer-grade hardware. Sharing weights guarantees that the same geometric patterns are encoded identically regardless of which branch processes them, and halves the number of trainable parameters.

3.6.1. Shared PointNet Encoder

Each input patch $P\in {\mathbb{R}}^{N\times 3}$ is first centroid-normalized by subtracting the patch mean. The normalized coordinates are processed by a PointNet encoder comprising (i) an input T-Net that learns a 3 × 3 spatial transformation to canonicalize the point cloud orientation; (ii) a shared Multi-Layer Perceptron (MLP) block (3→64→64); (iii) a feature T-Net that learns a 64 × 64 feature-space alignment; (iv) a second MLP block (64→128→D); and (v) global max-pooling that aggregates per-point features into a single D-dimensional descriptor. In the best-performing configuration, the embedding dimension is D = 1024. Both T-Nets are regularized by an orthogonal loss

$$L_{ortho}={‖\mathit{I}-{\mathit{A}\mathit{A}}^{\mathit{T}}‖}_F^2,$$

(5)

which encourages the learned transformations to remain close to rigid rotations.

3.6.2. Feature Fusion

Given the reference and current embeddings ${\mathit{f}}_{ref}, {\mathit{f}}_{cur}\in {\mathbb{R}}^D$, the fusion module computes two relational features: the absolute difference $\left|{\mathit{f}}_{ref}-{\mathit{f}}_{cur}\right|$ and the element-wise (Hadamard) product ${\mathit{f}}_{ref}\odot {\mathit{f}}_{cur}$. These are concatenated to form a 2D fused vector. The absolute difference captures the magnitude of per-dimension discrepancy, while the product encodes co-activation patterns; together they provide complementary views of the pairwise relationship.

3.6.3. Chamfer Statistics Branch

In addition to the learned PointNet features, an optional geometric branch computes Chamfer distance statistics directly from the raw point coordinates, providing an explicit low-level cue that complements the learned high-level representation. For each forward pass, n = 128 points are sampled uniformly at random without replacement from each input patch. During training, this sub-sampling is re-drawn stochastically; during evaluation, all random operations follow the fixed global seed reported in Section 3.8. Pairwise Euclidean distances are then computed using torch.cdist, and nearest-neighbor distance vectors are extracted in both directions (reference-to-current and current-to-reference). Bidirectionality is always included: four summary statistics are computed for each direction, specifically the mean, maximum, median, and 90th percentile, yielding an eight-dimensional descriptor in total. These values are transformed using log(1 + x) for numerical stability and passed through a two-layer MLP (8→64→32) to produce a 32-dimensional Chamfer feature vector. Gradients are not backpropagated through the Chamfer computation, since this branch is used only as a fixed auxiliary geometric prior rather than as a learned semantic representation. To reduce shortcut reliance on this handcrafted cue, stochastic gating zeros the Chamfer feature with probability p = 0.2 during training. This value was selected empirically as a moderate regularization setting that discourages over-reliance on the Chamfer branch while preserving the shared PointNet encoder as the primary learned representation. Because these statistics are derived solely from bidirectional nearest-neighbor geometry, they capture geometric discrepancy rather than high-level semantic attributes; consequently, the branch is intended to complement the learned PointNet representation, not to model appearance-only or texture-based changes.

3.6.4. Classification Head

The fused PointNet features (2 × 1024 = 2048 dimensions) and the Chamfer features (32 dimensions) are concatenated into a 2080-dimensional vector (when the Chamfer branch is active) or 2048-dimensional vector, which is passed through a three-layer classifier: Linear (2080→256)-ReLU-Dropout (0.5)-Linear (256→128)-ReLU-Dropout (0.5)-Linear (128→3). The output logits correspond to the three classes; at inference, the softmax probability of Class 1 serves as the change confidence score.

3.7. Loss Function

Training minimizes a composite objective comprising a classification term and an orthogonal regularization term:

$$L=L_{focal}+{\lambda }_{ortho}·L_{ortho}.$$

(6)

3.7.1. Focal Loss

The classification term employs focal loss [32], which modulates the standard cross-entropy by a factor ${(1-p_t)}^{\gamma }$ that down-weights well-classified (easy) examples and concentrates the gradient on hard, misclassified samples:

$$L_{focal}=-{\sum }_{c = 0}^2{{\alpha }_c(1-p_c)}^{\gamma }\mathit{log}\left(p_c\right),$$

(7)

where $p_c$ is the softmax probability assigned to the true class. The class-balancing weights $\alpha =[0.25, 0.35, 0.40]$ are set inversely proportional to the class sampling frequencies (50%, 35%, 15%), and the focusing parameter $\gamma =2.0$ provides strong down-weighting of easy examples. This choice was motivated by pilot experiments in which standard cross-entropy loss yielded high accuracy on the dominant no-change class but poor recall on the change class; focal loss substantially improved the precision–recall trade-off for Class 1.

3.7.2. Orthogonal Regularization

The T-Net regularization term averages the orthogonality violations across all four transformation matrices (input and feature T-Nets for both the reference and current branches):

$$L_{ortho}={\displaystyle \frac{1}{4}}{\sum }_k{‖\mathit{I}-{\mathit{A}}_{\mathit{k}}{\mathit{A}}_{\mathit{k}}^T‖}_F^2,$$

(8)

with ${\lambda }_{ortho}=0.001$. This regularization, originally proposed in [9], stabilizes the learned transformations and prevents degenerate solutions.

3.8. Training Procedure

The model is trained with the Adam optimizer (learning rate $\eta ={10}^{-3}$, weight decay ${10}^{-4}$) and a cosine annealing schedule that decays the learning rate to ${10}^{-5}$ over the course of training. The batch size is 16 and training runs for up to 1000 epochs, with early stopping triggered if the best validation metric does not improve for 100 consecutive epochs. Each training epoch samples 2000 pairs from the training split and 500 pairs from the validation split, with class ratios as specified in Section 3.5.3. The spatial tile set is divided into a 90%/10% train/validation split to ensure that validation pairs are drawn from held-out spatial regions.

Data augmentation is applied independently to each branch of each pair: Gaussian jitter ($\sigma =0.005 \mathrm{m}$), point dropout (5% of points dropped and replenished by nearest-neighbor duplication with ${\sigma }_{pad}=0.002$ m), and residual pose jitter (translation ${\sigma }_t=0.01$ m, yaw ${\sigma }_{\psi }=1°$). These augmentations simulate the sensor noise, occlusion variability, and registration residuals encountered in practice. All random seeds are fixed (seed = 42) for reproducibility.

The best checkpoint is selected by monitoring validation accuracy, prioritizing the minimum per-class precision to ensure balanced performance across all three classes. Training was conducted on a single NVIDIA GTX 1650 GPU (4 GB VRAM); a full training run to convergence (276 epochs until early stopping) required approximately 8 h.

3.9. Inference Pipeline

3.9.1. Current Map Construction

At inference time, one or more new LiDAR scans are merged and registered to the reference coordinate frame using the same point-to-plane ICP procedure described in Section 3.3.2 and [26], followed by voxel downsampling at 5 cm to match the reference map density.

3.9.2. Tile-Level Comparison

For each reference tile (i, j, k), the corresponding region of the current map is extracted. If the extracted region contains at least $N_{min}=512$ points, it is FPS-subsampled to $N=4096$ points (with a fixed seed for deterministic behavior) and centroid-normalized. The reference and current patches are forwarded through the trained Siamese network, which outputs Class 1 softmax probability ${\widehat{y}}_{ijk}\in \left[0,1\right]$ as the tile-level change confidence score.

3.9.3. Threshold Optimization

A global detection threshold $\tau$ is selected by sweeping over the range [0.1, 0.9] and maximizing the F1-score on a set of simulation experiments with known ground truth (described in Section 4). The optimal threshold $\tau =0.17$ was found to balance precision and recall for the range of anomaly types and magnitudes tested.

3.9.4. Change Localization

Tiles whose Class 1 probability exceeds the threshold (${\widehat{y}}_{ijk}>\tau$) are flagged as changed. The centers of flagged tiles are mapped back to 3D coordinates in the global reference frame, producing a spatial heatmap that localizes the detected changes within the environment. This heatmap can be overlaid on the reference map for visual inspection or consumed by downstream decision-making systems.

Because inference is performed independently for each tile, objects that cross tile boundaries may produce multiple adjacent positive tile predictions. In such cases, we interpret spatially contiguous activated tiles as one fragmented but coherent change region in the output heatmap. No explicit graph-based or connected-component aggregation is currently applied; however, such post-processing would be straightforward to incorporate and could improve the delineation of large objects spanning several neighboring tiles.

4. Experimental Setup

The experimental protocol is designed to evaluate three aspects of the proposed framework: detection performance, robustness to controlled change conditions, and practical deployment behavior under lightweight hardware constraints.

4.1. Dataset Description

The dataset comprises 42 LiDAR scans (25 scans were used to create the global map) collected by a Unitree Go2 quadruped robot traversing an indoor corridor environment, as described in Section 3.3.1. After multi-scan registration and voxel downsampling at 5 cm [26], the resulting global reference map is partitioned into spatial tiles of $1.5\times 1.5\times 1.0$ m. Tiles containing fewer than 512 points are discarded. Each retained tile is subsampled to n = 4096 points via FPS, producing the set of reference patches used for both training and inference. Additionally, the dense (pre-FPS) point sets for each tile are stored separately to support the pre-FPS anomaly injection strategy described in Section 3.5.2. The spatial tile set is split 90%/10% into training and validation subsets, with the split performed over tile identifiers to ensure that validation pairs are drawn from held-out spatial regions. This spatial split is fixed across runs, and all random operations in training and evaluation use the same global seed (seed = 42), ensuring reproducibility of the reported protocol.

Table 3. LiDAR dataset statistics and patch sampling parameters.

Property	Value
Number of raw scans	42
Sensor	Unitree 4D LiDAR L1
Voxel resolution	5 cm
Tile dimensions	1.5 × 1.5 × 1.0 m
Points per patch (N)	4096
Minimum points per tile	512
Train/validation split	90%/10% (spatial)
Training pairs per epoch	2000
Validation pairs per epoch	500

summarizes the dataset statistics.

4.2. Simulation Experiment Design

Because the target environment lacks labeled real-world changes, we evaluate the framework through controlled simulation experiments in which synthetic anomalies of known type, magnitude, and spatial extent are injected into a subset of the original scans before map construction. This approach provides exact ground truth for quantitative evaluation while simulating realistic deployment conditions: the robot collects scans, some of which contain changes; a current map is built from all scans (so anomalies are voxelized into the map); and the model must detect which tiles have changed relative to the reference.

Each experiment follows a standardized protocol:

• All 25 original scans (map scans) are copied to an experiment directory;
• Anomalies are injected into the selected scans at the specified tiles using the operations described in Section 3.5.1;
• A current map is built from all scans (including modified ones) via ICP registration and voxel downsampling at 5 cm;
• Tile-level inference is performed against the stored reference patches;
• Predictions are evaluated against the known ground truth.

The experiments are organized into four groups that systematically vary one parameter at a time while holding the others fixed, plus a baseline and two challenge conditions. The full experiment matrix is presented in

Table 4. Simulation experiment matrix. All experiments use eight tiles per modified scan unless otherwise noted.

Experiment	Change Type	Magnitude (cm)	Density (%)	Modified Scans	Tiles/Scan
Magnitude series
magnitude_5 cm	add	5	30	3	8
magnitude_10 cm	add	10	30	3	8
magnitude_15 cm	add	15	30	3	8
magnitude_20 cm	add	20	30	3	8
magnitude_30 cm	add	30	30	3	8
Density series
density_10 pct	add	15	10	3	8
density_30 pct	add	15	30	3	8
density_50 pct	add	15	50	3	8
Type series
type_add	add	15	35	3	8
type_remove	remove	15	35	3	8
type_move	move	15	35	3	8
Challenge
challenge_easy	add	30	50	4	8
challenge_hard	add	5	10	2	1
Baseline
baseline_no_change	-	-	-	-	-

.

4.2.1. Change Magnitude Experiments

Five experiments evaluate the model’s sensitivity to the spatial magnitude of injected changes, ranging from 5 cm (near the noise floor of the 5 cm voxel grid) to 30 cm (a clearly visible displacement). All magnitude experiments use the add operation with 30% point density, applied to eight tiles in each of three modified scans. This series tests whether the model can detect subtle changes that may be partially absorbed by voxel downsampling and identifies the minimum detectable change magnitude for the system.

4.2.2. Change Density Experiments

Three experiments vary the fraction of tile points affected by the anomaly: 10%, 30%, and 50%. All use the add operation at 15 cm magnitude across eight tiles per scan in three scans. This series evaluates how the number of displaced points within a tile affects detection performance, simulating the difference between a small object appearing (sparse change) versus a large-scale rearrangement (dense change).

4.2.3. Change Type Experiments

Three experiments isolate each anomaly operation: add, remove, and move at a fixed magnitude of 15 cm and 35% density, applied to eight tiles per scan in three scans. The add experiment simulates the introduction of new objects, the remove experiment simulates object disappearance, and the move experiment simulates object displacement. This series reveals whether the model exhibits differential sensitivity to the nature of geometric change.

4.2.4. Challenge Experiments

Two boundary condition experiments test the extremes of the model:

• Easy challenge: Large magnitude (30 cm), high density (50%), four modified scans with eight tiles each. This represents clearly visible, widespread changes that the model should detect with high confidence.
• Hard challenge: Small magnitude (5 cm), low density (10%), two modified scans with only one tile each. This represents subtle, localized changes at the detection limit of the system.

4.2.5. Baseline (No Change)

A baseline experiment uses the full set of 25 original scans without any modification. The current map is constructed identically to the reference map. In this condition, no tiles should be flagged as changed; any detections are false positives. This experiment provides a direct measurement of the system’s false positive rate under normal operating conditions.

4.3. Evaluation Metrics

All experiments are evaluated at the tile level. For each experiment, the set of tiles is partitioned into positives (tiles known to have been modified) and negatives (all remaining tiles). Each tile’s Class 1 softmax probability is compared against the detection threshold τ to produce a binary prediction.

From the resulting confusion matrix, we compute the standard classification metrics:

$$Precision={\displaystyle \frac{TP}{TP+FP}},$$

(9)

$$Recall={\displaystyle \frac{TP}{TP+FN}},$$

(10)

$$F_1={\displaystyle \frac{2·Precision·Recall}{Precision+Recall}},$$

(11)

$$Accuracy={\displaystyle \frac{TP+TN}{TP+FP+FN+TN}},$$

(12)

where TP, FP, FN, and TN denote the counts of true positives, false positives, false negatives, and true negatives, respectively. We additionally report the true positive rate (TPR = Recall) and false positive rate (FPR = FP/(FP + TN)) at each threshold to construct Receiver Operating Characteristic (ROC) curves across the threshold range [0.1, 0.9].

The global detection threshold τ = 0.17 was obtained by sweeping candidate thresholds on the simulation benchmark used in this study and selecting the value that maximized the aggregated F1-score. The selected threshold was then fixed for the reported thresholded analyses. To complement the selected operating threshold, additional score-based analyses are provided to characterize the detector beyond a single binary operating point. These include a threshold-sensitivity analysis, a precision–recall curve, and a threshold-sensitivity analysis stratified by anomaly type. Together, these visualizations provide a more complete picture of score separation, operating-point behavior, and the observed asymmetry between anomaly types on the simulation benchmark.

Figure 5 shows the sensitivity of precision, recall, and F1-score to the detection threshold τ. The selected threshold τ = 0.17 lies near the pooled F1-optimal operating region while reflecting the expected trade-off between recall at lower thresholds and precision at higher thresholds.

Figure 6 presents the precision–recall curve for tile-level change detection across the simulation benchmark. The marked operating point at τ = 0.17 indicates where the thresholded evaluation is performed, while the overall curve provides a threshold-independent view of the model’s ranking behavior under class imbalance.

Figure 7 further decomposes threshold sensitivity by anomaly type. The add and move conditions remain comparatively stable across a broad threshold range, whereas the remove condition is consistently more sensitive and achieves a lower peak F1-score. This behavior is consistent with the weaker geometric evidence produced by local point removal and motivates reporting threshold behavior separately by anomaly type.

5. Results

This section presents the experimental results obtained from the trained Siamese PointNet model across all 14 simulation experiments described in Section 4.2. We first report training convergence behavior, then provide an overview of detection performance through the main results table, followed by a detailed analysis of each experimental series. We conclude with an embedding-space analysis and a qualitative spatial localization example.

5.1. Training Convergence

Figure 8 presents the training and validation loss and accuracy curves over the full training run. The focal loss decreases rapidly during the first 20 epochs and stabilizes below 0.02 by epoch 100, with no visible divergence between training and validation losses thereafter. This close tracking indicates that the model does not overfit the synthetically generated training pairs, despite the relatively small spatial extent of the corridor environment.

Validation accuracy rises from approximately 50%, consistent with random chance for a balanced three-class problem, to over 90% by epoch 60, and plateaus around 96% for the final 150 epochs. The per-class validation accuracies confirm that all three classes (no change, changed, and different tile) are learned effectively. Training was terminated by the early-stopping criterion (patience = 100 epochs) after 276 epochs, with the best checkpoint selected by maximum validation accuracy.

5.2. Overall Detection Performance

Table 5. Tile-level change detection results for all 14 simulation experiments at the optimized detection threshold τ = 0.17. TP: true positives; FP: false positives; FN: false negatives; TN: true negatives. The aggregated row pools confusion matrix counts across all experiments.

Experiment	Type	Mag. (cm)	Dens. (%)	TP	FP	FN	TN	Precision	Recall	F1	Accuracy
Magnitude series
magnitude_5 cm	add	5	30	18	1	6	147	0.95	0.75	0.84	0.96
magnitude_10 cm	add	10	30	19	5	4	144	0.79	0.83	0.81	0.95
magnitude_15 cm	add	15	30	19	5	2	146	0.79	0.90	0.84	0.96
magnitude_20 cm	add	20	30	19	1	1	151	0.95	0.95	0.95	0.99
magnitude_30 cm	add	30	30	20	3	4	145	0.87	0.83	0.85	0.96
Density series
density_10 pct	add	15	10	17	2	7	146	0.89	0.71	0.79	0.95
density_30 pct	add	15	30	21	2	3	146	0.91	0.88	0.89	0.97
density_50 pct	add	15	50	18	1	6	147	0.95	0.75	0.84	0.96
Type series
type_add	add	15	35	20	1	2	149	0.95	0.91	0.93	0.98
type_remove	remove	15	35	12	7	8	145	0.63	0.60	0.62	0.91
type_move	move	15	35	19	4	5	144	0.83	0.79	0.81	0.95
Challenge
challenge_easy	add	30	50	29	5	3	135	0.85	0.91	0.88	0.95
challenge_hard	add	5	10	1	1	1	169	0.50	0.50	0.50	0.99
Baseline
baseline_no_change	-	-	-	0	0	0	172	-	-	-	1.00
Aggregated				232	38	52	2086	0.86	0.82	0.84	0.96

summarizes the tile-level detection results for all 14 simulation experiments at the fixed operating threshold τ = 0.17. For a top-level overview, the key reference points are the Aggregated row (bottom of Table 5: P = 0.86, R = 0.82, F1 = 0.84) and the per-series weak point: the remove-type anomaly (F1 = 0.62) and the hard challenge (F1 = 0.50) define the system’s current detection floor, while the magnitude and type-add experiments (F1 = 0.84–0.95) define its ceiling. Across the 12 experiments containing synthetic changes, the model achieved mean F1 = 0.85, with per-experiment values ranging from 0.62 to 0.95. Precision remained high across the standard experiments, indicating that most detections were not driven by unstable false-positive behavior.

When confusion-matrix counts are aggregated across all 14 experiments, the model yields TP = 232, FP = 38, FN = 52, and TN = 2086, corresponding to Precision = 0.86, Recall = 0.82, F1 = 0.84, and Accuracy = 0.96. This indicates that the dominant residual error is missed changed tiles rather than excessive false alarms, which is consistent with the weaker response observed later for removal-type anomalies.

5.3. Benchmark Comparison with Geometric and Learned Baselines

To strengthen the quantitative validation, we benchmarked the proposed SPICD-Net against two classical geometric baselines, C2C and M3C2-approx, and two learned Siamese baselines based on PointNet++ and KPConv under the same indoor tile-level protocol. All methods were evaluated on the same 14 simulation experiments, using the same reference tiling, current-map construction procedure, changed-tile ground truth, and micro-aggregated metrics.

Table 6. Benchmark comparison under the unified indoor tile-level protocol. All methods were evaluated on 14 simulation experiments with the same reference tiling, current-map construction, and changed-tile ground truth. Values are micro-aggregated at the best threshold per method. M3C2-approx is a PCA-normal point-to-plane approximation, not the official CloudCompare implementation. Inference time: 172 tiles, single NVIDIA GTX 1650 (4 GB VRAM).

Method	Best τ	Precision	Recall	F1	Accuracy	TP	FP	FN	TN	Mean Infer. (s)
SPICD-Net (ours)	0.17	0.859	0.817	0.838	0.963	232	38	52	2086	22.4
Siamese-PointNet++	0.20	0.930	0.845	0.886	0.974	240	18	44	2106	131.1
Siamese-KPConv	0.40	0.564	0.849	0.678	0.905	241	186	43	1938	126.1
C2C	0.01	0.972	0.366	0.532	0.924	104	3	180	2121	25.8
M3C2-approx	0.03	0.000	0.000	0.000	0.882	0	0	284	2124	23.1

summarizes the best threshold selected for each method together with Precision, Recall, F1-score, Accuracy, confusion-matrix counts, and mean inference time.

Among the evaluated methods, Siamese-PointNet++ achieved the highest raw detection performance under the unified indoor tile-level protocol, reaching Precision = 0.930, Recall = 0.845, F1 = 0.886, and Accuracy = 0.974. However, this improvement came at a substantially higher computational cost: 131.1 s vs. 22.4 s for SPICD-Net, a 5.8× speedup. Siamese-KPConv similarly required 126.1 s, a 5.6× overhead. For broader context, the leading methods in the broader literature (PBFormer [6], PGN3DCD [5]) were reported on server-class Tesla V100 GPUs; therefore, those runtimes are not directly comparable to the consumer-GPU measurements reported in Table 6. The proposed SPICD-Net achieved Precision = 0.859, Recall = 0.817, F1 = 0.838, and Accuracy = 0.963, indicating competitive detection quality at much lower deployment cost. Under the present training and evaluation budget, the Siamese-KPConv baseline did not reach comparable performance (F1 = 0.678), while the classical C2C baseline behaved conservatively, attaining very high precision (0.972) but low recall (0.366). The M3C2-approx baseline failed to recover changed tiles under this merged-map tile-level protocol, suggesting that this approximation is not well-suited to the present setting. Overall, these results position SPICD-Net not as the highest-scoring model in absolute F1, but as the most computationally efficient learned alternative tested, achieving competitive detection quality at 5.8× lower inference cost than the next-best learned method on the same consumer GPU.

5.4. Change Magnitude Sensitivity

The magnitude series (5–30 cm) evaluates detection performance as spatial displacement increases while anomaly type (add) and density (30%) remain fixed. The model achieves F1 ≥ 0.81 across all five magnitudes, indicating sensitivity even at 5 cm, which matches the voxel-grid resolution.

Performance peaks at 20 cm (F1 = 0.95, P = 0.95, R = 0.95), with only one false positive and one false negative. At 5 cm, precision remains high (0.95), but recall drops to 0.75, indicating that several changed tiles remain too subtle after voxel downsampling. The 10 cm condition yields a lower F1 (0.81) mainly because of increased false positives, while performance at 30 cm (F1 = 0.85) declines slightly because of both additional false positives and false negatives, consistent with a stronger interaction between large displacements and the boundary between the “changed” and “different tile” classes.

Overall, the model remains effective across the full tested displacement range, with the strongest performance around 20 cm.

5.5. Change Density Effects

The density series varies the fraction of tile points affected by the anomaly (10%, 30%, 50%) at fixed magnitude (15 cm) and anomaly type (add). The relationship between density and F1 is non-monotonic: density 30 pct achieves the highest F1 (0.89), followed by density 50 pct (0.84) and density 10 pct (0.79).

At low density (10%), the model produces seven false negatives because only about 410 of the 4096 points are displaced, yielding a weak geometric signal that is often indistinguishable from normal FPS variation. At high density (50%), precision remains high (0.95), but recall drops to 0.75 with six false negatives. This suggests that very dense perturbations may alter the tile geometry enough to weaken the distinction between a locally changed tile and a geometrically inconsistent pair. The 30% condition provides the best balance between detectability and preservation of tile identity.

5.6. Change Type Analysis

The type series isolates each anomaly operation (add, remove, move) at a fixed magnitude (15 cm) and density (35%). Figure 9 presents the confusion matrices for this series, revealing a pronounced asymmetry in detection performance across change types.

The add operation achieves the highest F1-score (0.93), with only one false positive and two false negatives. This is consistent with the fact that point addition introduces new local geometric structures and density increases, producing a strong signal in both the embedding-difference and Chamfer branches.

The move operation yields intermediate performance (F1 = 0.81), with four false positives and five false negatives. Because moving points creates both a local deficit and a local surplus, it produces a mixed geometric signal that is less separable than pure addition.

The remove operation exhibits the weakest performance (F1 = 0.62), with seven false positives and eight false negatives, the highest error counts among the standard experiments. Point removal is intrinsically harder to detect because it reduces local support without introducing new structures, and this signal is further weakened after FPS resampling redistributes the remaining points. In addition, Chamfer sensitivity is asymmetric in this case: the missing region is more strongly reflected from the reference-to-current direction than from the current-to-reference direction. Together, these effects explain the clear performance gap between remove and the other anomaly types.

5.7. Boundary Conditions and Baseline

Figure 10 presents the confusion matrices for the baseline and two challenge experiments, which test the system’s behavior at the extremes of its operating range.

The baseline (no-change) experiment reconstructs the current map from the original 25 scans without modification. All 172 tiles are correctly classified as unchanged (172 TN, 0 FP), yielding a false positive rate of zero. This confirms that natural variation from ICP registration and voxel downsampling alone does not trigger false alarms at the operating threshold τ = 0.17.

The easy challenge (30 cm magnitude, 50% density, four modified scans) achieves F1 = 0.88, with 29 of 32 changed tiles correctly detected. The remaining three false negatives and five false positives indicate that tile-level detection remains imperfect even under favorable conditions, likely because anomaly visibility depends on local tile geometry.

The hard challenge (5 cm magnitude, 10% density, two modified scans with one tile each) represents the practical detection floor. Of the two changed tiles, one is detected and one is missed, with one false positive, yielding F1 = 0.50. This is consistent with a signal near the voxel-resolution limit. Despite the low F1, the system still maintains 99% accuracy (169/170 unchanged tiles correct), showing that minimal-change detection does not lead to widespread false alarms.

5.8. Robustness to Noise and Registration Perturbations

To complement the controlled change-magnitude, density, and anomaly-type experiments, we further evaluated the robustness of the trained SPICD-Net under explicit inference-time perturbations. Two noise families were considered at the patch level, namely Gaussian coordinate jitter and random point dropout, while two registration-related perturbation families were applied at the map level, namely translational misalignment and yaw rotation error. Figure 11 summarizes the resulting Precision, Recall, F1-score, and false positive rate (FPR) under increasing perturbation levels, evaluated on a representative subset of benchmark experiments.

The robustness analysis reveals a strongly asymmetric behavior across perturbation types. SPICD-Net remains comparatively tolerant to moderate point dropout, with F1 decreasing from 0.832 at baseline to 0.534 at 30% dropout, although this degradation is accompanied by a steady rise in false positives. By contrast, Gaussian jitter produces a rapid collapse in precision and a corresponding increase in FPR, indicating that small coordinate perturbations can destabilize the geometric comparison signal used by the model. The strongest degradation is observed under explicit registration perturbations: both translational offsets and yaw errors sharply reduce F1 and drive FPR upward, showing that the current framework is highly sensitive to misalignment between the reference and current maps. These results indicate that SPICD-Net is substantially more tolerant to missing points than to geometric noise or registration error, and they support the conclusion that accurate upstream registration remains critical for reliable deployment.

5.9. Embedding Space Analysis

To understand the model’s internal representations, we perform a Principal Component Analysis (PCA) projection of the fused feature vectors (concatenation of element-wise absolute difference and element-wise product of the Siamese branches) computed across all experiments. Figure 12 visualizes these embeddings colored by classification outcome (TP, FP, FN, TN).

The projection reveals a clear structure in the learned feature space: true negatives (gray) cluster tightly at low PC1 values, whereas true positives (green) extend across a broad arc toward much larger PC1 values. This indicates that the fused embedding captures a meaningful notion of geometric dissimilarity between unchanged and changed tile pairs.

False negatives (blue squares) are concentrated in the transition region between the TN and TP clusters, indicating that missed detections arise when the change signal remains close to the decision boundary. By contrast, false positives (red triangles) lie mainly along the periphery of the TN cluster, suggesting that they are associated with tiles whose reference geometry is intrinsically less stable.

The L2 distance between reference and current embeddings increases with predicted change probability, confirming that the classifier output is consistent with the underlying embedding geometry.

5.10. Spatial Localization

Figure 13 overlays the tile-level predictions on the original corridor map for the magnitude_15 cm experiment, providing a qualitative assessment of spatial localization accuracy. True positive tiles (green) are correctly identified in the upper portion of the map where anomalies were injected, while the majority of the map remains correctly classified as unchanged (gray). The two false negative tiles (blue) are spatially adjacent to correctly detected tiles, suggesting that the model’s miss is a matter of threshold proximity rather than a systematic spatial blind spot. The five false positive tiles (red) are distributed across the map without obvious spatial clustering, indicating that they arise from individual tile properties (e.g., edge effects or low point density in peripheral tiles) rather than from a systematic spatial bias in the model.

Figure 14 illustrates a representative failure case from the remove-type experiment. In this example, the removed structure produces a predicted change score of 0.161, which falls just below the operating threshold τ = 0.17, resulting in a false negative. This behavior is consistent with the analysis in Section 5.6, where removal-type anomalies yielded the weakest performance because local deletions reduce occupancy without introducing a new geometric structure, and FPS resampling partially redistributes the remaining points.

Figure 15 presents the distribution of predicted Class-1 change probabilities for all changed and unchanged tiles aggregated across experiments. Most unchanged tiles concentrate below the threshold, while changed tiles occupy a broader range extending to high-confidence values. The overlap region near the decision boundary explains the remaining false positives and false negatives and provides a visual justification for the use of the global threshold τ = 0.17 introduced in Section 4.3.

5.11. Small-Scale Real-World Validation

To complement the synthetic benchmark and the controlled robustness analysis, we conducted a small-scale real-world validation on paired indoor LiDAR scans acquired in a different room that was not used during training. The pre-change set comprised 2 scans of the empty room, and the post-change set comprised 2 scans of the same room after a box was manually inserted into the scene, giving 4 scans in total. The scans were registered into a common coordinate frame using the same point-to-plane ICP pipeline described in Section 3.3.2, with identical quality thresholds (fitness > 0.55, RMSE < 0.05 m) and voxel downsampling at 5 cm resolution. The resulting reference and current maps were partitioned into spatial tiles of 1.5 × 1.5 × 1.0 m, retaining tiles with at least 512 points and subsampling each to n = 4096 points via FPS, yielding 67 tiles in total. No retraining or fine-tuning was performed; the model trained on the corridor environment was applied directly. Since the unseen room environment differs geometrically from the training corridor, the detection threshold was re-optimized on the real-world scan pair by sweeping τ over [0.1, 0.9] and selecting the value that maximized F1-score, yielding τ = 0.50 for this experiment.

Figure 16 illustrates the resulting tile-level detections in top and side views. The predicted changed region overlaps the manually altered area, supporting the conclusion that the proposed framework can transfer beyond the synthetic generation pipeline and detect genuine real-world scene changes in an unseen indoor environment.

The tile-level evaluation for this experiment yielded TP = 7, FP = 5, FN = 0, TN = 55, corresponding to Precision = 0.583, Recall = 1.000, F1-score = 0.737, and Accuracy = 0.925. The model successfully detected all tiles containing the inserted object (zero false negatives), while five neighboring tiles were incorrectly flagged as changed. The elevated false-positive count relative to the main synthetic benchmark is consistent with the increased geometric variability of an unseen room environment and the higher operating threshold required to suppress false alarms in an unseen domain.

This validation is limited to a small number of scans acquired in a single unseen room and should be interpreted as feasibility-oriented evidence rather than as a comprehensive real-world benchmark.

6. Discussion

6.1. Interpretation of Key Findings

Four findings from the simulation benchmark merit particular discussion.

First, detecting changes at the voxel-resolution floor (5 cm magnitude) indicates that even substantially attenuated geometric signals remain recoverable through the learned feature space. The embedding analysis (Section 5.9) supports this: changed tiles at 5 cm are displaced from the no-change cluster in PCA space, and this sensitivity is attributed to the pre-FPS injection strategy (Section 3.5.2), which exposes the model to the same voxelization artifacts at training time that it encounters at inference.

Second, the pronounced asymmetry between add-type (F1 = 0.93) and remove-type (F1 = 0.62) changes reveals a fundamental property of fixed-size point cloud comparison. When points are added, both the PointNet embedding and the Chamfer statistics register a clear signal: new geometric structures appear in the current cloud with no correspondence in the reference. When points are removed, FPS resampling redistributes the surviving points to maintain n = 4096, partially filling the gap and masking the change signal. This masking effect is compounded by the global max-pooling operation in PointNet [9], which is inherently less sensitive to local point absence than to local point addition. The finding aligns with observations from de Gélis et al. [3,4], who reported that removal-type changes consistently achieved lower per-class IoU than addition-type changes across multiple outdoor datasets, suggesting this asymmetry is a broader challenge in Siamese point cloud change detection rather than a property specific to this architecture.

Third, the zero false-positive rate in the no-change baseline experiment (172/172 tiles correct) is a critical result for practical deployment. In an autonomous inspection scenario, false alarms trigger unnecessary maintenance actions and erode operator trust. This result confirms that natural variation from ICP registration residuals and FPS stochasticity alone does not cross the detection threshold τ = 0.17, providing a reliable operating floor for the system.

Fourth, the robustness analysis (Section 5.8) reveals an asymmetric sensitivity profile: SPICD-Net tolerates moderate point dropout (F1 = 0.534 at 30% dropout) but degrades rapidly under Gaussian coordinate jitter and registration perturbations, with yaw rotation error producing the sharpest F1 decline and highest false-positive rate. This confirms that noise and misregistration are the primary operational risk factors, and that accurate ICP-based registration in Stage 1 is a prerequisite for reliable downstream change detection rather than an optional preprocessing step.

6.2. Design Choice Analysis

Three-class formulation. The inclusion of a dedicated “different tile” class (Class 2) as a hard negative during training (Section 3.1) forces the encoder to learn the distinction between geometric change within a spatial location and geometric dissimilarity between unrelated locations. This design prevents the model from misclassifying structurally different tiles as changed, as evidenced by the zero false-positive baseline and the consistently low FPR (1.6%) across all experiments.

Pre-FPS anomaly injection. Injecting anomalies into dense (pre-FPS) tile point clouds before subsampling to n = 4096 (Section 3.5.2) reduces the mismatch between synthetic training pairs and inference-time preprocessing. The magnitude series results support this design: the model detects 5 cm changes despite their substantial attenuation by 5 cm voxel downsampling, indicating that training on attenuated signals is effective.

Low detection threshold (τ = 0.17). The optimal threshold found by F1-maximization is notably low relative to the typical 0.5 default. This reflects the class imbalance inherent in the inference setting: only ~14% of tiles are changed in most experiments. A low threshold increases recall at a modest cost to precision, appropriate for inspection tasks where missing a genuine change is costlier than investigating a false alarm.

Focal loss: Using focal loss [32] with γ = 2.0 down-weights well-classified examples and redirects training toward harder, changed, and different-tile examples, substantially improving the precision–recall trade-off. The resulting model achieves balanced per-class validation accuracy of ~96%, confirming that the loss function effectively mitigates class imbalance.

6.3. Limitations

Several limitations should be acknowledged to contextualize the results and guide future work:

• Synthetic-only ground truth. Both training and the primary evaluation rely on synthetically generated geometric changes, because matched indoor LiDAR datasets with tile-level real change annotations are currently unavailable. The real-world validation in Section 5.11 further indicated that the detection threshold optimized on the synthetic benchmark (τ = 0.17) does not transfer directly to an unseen environment, where re-optimization yielded τ = 0.50, suggesting that environment-specific threshold calibration may be needed in practice. Generalization claims are therefore explicitly bounded to the tested single-corridor and single-room indoor settings.
• Sensitivity to noise, misregistration, and scene variability. The robustness analysis (Section 5.8) demonstrates that SPICD-Net tolerates moderate point dropout (F1 degrades from 0.832 at baseline to 0.534 at 30% dropout) but is substantially more sensitive to geometric noise and registration perturbations: Gaussian coordinate jitter produces a rapid collapse in precision, while translational and yaw misalignment sharply increase the false positive rate. This indicates that sensor noise and imperfect scan registration are the primary operational risk factors, and that the Stage 1 ICP pipeline quality directly constrains downstream detection reliability. Regarding variability across scenes, the single cross-scene transfer experiment (Section 5.11) showed that the model generalizes geometrically, achieving Recall = 1.000 in an unseen room, but at the cost of elevated false positives (Precision = 0.583) and a required threshold shift from τ = 0.17 to τ = 0.50, indicating sensitivity to inter-scene geometric distribution differences. Broader evaluation across multiple environments with varying geometry, clutter, and LiDAR placement remains an open direction for future work.
• Tile-level granularity. The framework produces binary decisions at the tile-level (1.5 × 1.5 × 1.0 m), limiting spatial precision. Finer-grained localization through overlapping tiles, adaptive tiling, or point-level refinement could improve localization.
• Remove operation sensitivity. Removal exhibits the weakest performance (F1 = 0.62), likely due to the global max-pooling and fixed-size resampling. Architectural changes targeting removal detection would be valuable.
• Geometric rather than semantic sensitivity. The framework relies on raw LiDAR coordinates and does not incorporate RGB, intensity, or semantic labels, making it primarily sensitive to geometric changes.

6.4. Future Work

Several directions emerge from the current findings:

• Real-world validation. Evaluate the framework on additional real physical scene changes, either through newly collected indoor LiDAR sequences with staged modifications or through compatible public multi-temporal datasets, to test generalization beyond the present single-environment benchmark.
• Backbone upgrade. Replacing the PointNet encoder with PointNet++ [10] or KPConv [3] to introduce hierarchical local feature extraction while retaining the Siamese training framework.
• Point-level change segmentation. Extend the framework from tile-level classification to point-level change segmentation by replacing the classification head with a segmentation decoder.
• Temporal monitoring. Extend the framework to process temporal sequences, enabling continuous monitoring and distinction between transient and persistent changes.
• Adaptive thresholding. Develop tile-specific or context-adaptive thresholds based on local point density or geometric complexity.

7. Conclusions

This paper presented SPICD-Net, a lightweight Siamese PointNet framework for autonomous indoor change detection in 3D LiDAR point clouds. The proposed pipeline integrates automated multi-scan registration, spatial tiling, synthetic anomaly generation, and tile-level Siamese classification into a complete end-to-end workflow that can be deployed on a quadruped robotic platform. By generating synthetic add, remove, and move anomalies directly from reference tiles, the method avoids manual point-level annotation while preserving controlled supervision over change magnitude, density, and type. The three-class formulation, distinguishing no change, changed, and different tile, further improves robustness by teaching the network to separate genuine local changes from geometrically inconsistent comparisons.

Systematic evaluation on the indoor corridor dataset showed that the framework achieves strong and stable tile-level performance across controlled simulation experiments, with high precision, recall, and F1-score over a broad range of anomaly conditions. The results also revealed a consistent asymmetry between addition- and removal-type changes: additions were detected more reliably, whereas removals remained more difficult, particularly when only a small number of points was affected after fixed-size sampling. This behavior highlights an important representational limitation of fixed-cardinality point cloud processing and suggests that subtle subtractive changes remain a challenging case for lightweight Siamese architectures.

The benchmark comparison further clarified the practical contribution of the proposed method. Among the learned baselines, Siamese-PointNet++ achieved the highest raw detection accuracy, while SPICD-Net provided the most favorable trade-off between detection quality and computational efficiency on consumer-grade hardware. In particular, SPICD-Net preserved competitive F1-score and accuracy while requiring substantially less inference time than heavier learned alternatives, supporting its suitability for resource-constrained robotic deployment. Classical geometric baselines also remained informative: C2C acted as a conservative high-precision baseline, whereas the tested M3C2-approx configuration was less effective under the present merged-map tile-level protocol.

Additional robustness experiments showed that the framework is relatively tolerant to moderate point dropout, but more sensitive to geometric noise and registration perturbations. Performance degraded progressively under Gaussian jitter, translational misalignment, and especially yaw rotation error, indicating that accurate map registration remains a critical prerequisite for reliable downstream change detection. These findings are consistent with the geometry-driven nature of the method and help define the operational limits of the proposed pipeline in practical robotic settings.

A small real-world validation in an unseen room further demonstrated the feasibility of transferring the trained model beyond the simulation setup. Although this experiment was intentionally limited in scale and should be interpreted as preliminary evidence rather than a full benchmark, the model successfully detected introduced real physical changes after registration and tile-based inference, supporting the practical relevance of the proposed training strategy.

Overall, the study shows that a lightweight Siamese PointNet architecture trained entirely on synthetic anomalies can provide effective indoor 3D change detection without manual annotation and with computational requirements compatible with low-cost robotic hardware. At the same time, the experiments identify several directions for future work: improving sensitivity to removal-type changes, increasing robustness to registration residuals, extending validation on larger real-world datasets, and exploring hybrid formulations that combine the efficiency of SPICD-Net with stronger local geometric modeling.