Non-contact three-dimensional reconstruction of thin, sheet-like heritage artifacts poses significant geometric and registration challenges. Due to their fragility, these artifacts cannot be suspended or equipped with artificial markers, necessitating independent acquisition of their front and back surfaces. Subsequent registration proves difficult due to the limited number of shared geometric features and the scarcity of explicit physical constraints, which may result in rotational ambiguity, instability, and structural collapse during iterative optimization. To address these challenges, we propose a geometry-constrained bidirectional point cloud registration method specifically tailored for thin, sheet-like heritage artifacts. The method integrates semantic-guided preprocessing, Principal Component Analysis (PCA)-based geometric normalization, and a thickness-aware registration strategy. The estimated physical thickness is incorporated as a geometric constraint to preserve structural integrity during registration. Rotational ambiguity is resolved by evaluating a finite set of global rotation hypotheses, each refined using the point-to-plane Iterative Closest Point (ICP) algorithm, with the optimal transformation selected via a geometry-aware fitness criterion consistent with the thickness scale. Experimental results show that the proposed method achieves competitive or improved performance in most cases, particularly in projected area consistency and physically plausible front-back alignment. In addition, the thickness-aware constraint and rotation hypothesis evaluation reduce the risk of degenerate configurations in which the two surfaces are incorrectly flipped while still yielding deceptively acceptable numerical scores, supporting reliable non-contact digitization of delicate and thin heritage artifacts. Implementation details are available at https://zyz-nwpu.github.io/GCBPCR/.
Figures & tables
Figure 1. Overview of the proposed geometry-constrained bidirectional point cloud registration pipeline for thin sheet-like artifacts. The workflow begins with independent front- and back-side multi-view reconstruction using an SfM–MVS pipeline. Semantic-guided purification is then applied to remove background structures and reconstruction artifacts, producing artifact-only point clouds for both sides. Next, PCA-based canonical normalization aligns each point cloud to its intrinsic geometric axes and enables robust estimation of the artifact thickness. Based on this canonical representation, a geometry-constrained registration stage evaluates a finite set of rotation hypotheses and refines each candidate using point-to-plane ICP initialized with a thickness-aware offset. The optimal transformation is selected using a thickness-aware fitness criterion, and the aligned point clouds are finally merged to obtain the complete artifact model. A flowchart of the complete registration pipeline. Double-sided image acquisition and semantic-guided purification produce separate artifact-only point clouds for the front and back surfaces. PCA aligns each surface with three principal axes, with the smallest-eigenvalue axis representing thickness. Thickness estimation generates several rotation hypotheses, which pass through coarse registration, point-to-plane ICP refinement, and fitness evaluation. The highest-scoring hypothesis produces the final merged artifact model.
Figure 2. Artifact-aware multi-view reconstruction and purification pipeline. The front and back surfaces are digitized through double-sided independent multi-view acquisition. The captured image sets are processed using an SfM–MVS reconstruction framework to generate dense point clouds. Semantic masks are predicted for each image and integrated with 3D structural filtering to refine the reconstructed data, yielding artifact-only point clouds for subsequent registration. A two-branch purification workflow beginning with front-side and back-side image sets. The upper branch applies Structure from Motion and Multi-View Stereo to estimate camera poses and dense point clouds. The lower branch applies Segment Anything Model 2 to generate binary artifact masks. Camera information, dense point clouds, and masks converge at mask-guided point-cloud purification, producing clean and separate point clouds for the two artifact surfaces.
Figure 3. Visualization of point cloud registration under different hypotheses. The first column shows the direct registration results of the two point clouds. The second column presents an exploded view where the two point clouds are spatially separated with a large offset to better observe their relative rotations. The best alignment hypothesis is highlighted. Four registration hypotheses for the two surfaces of an elongated decorated artifact are compared in rows: identity, flip about the X axis, flip about the Y axis, and flip about the Z axis. The left column shows the merged registration result, while the right column separates the two surfaces vertically for inspection. The flip-Y hypothesis produces the most consistent orientation and boundary correspondence and is identified as the best result.
Figure 4. Front and back views of all artifacts involved in the experiments. Front and back photographs of seven thin cultural-heritage artifacts arranged as labeled pairs. The collection includes small teardrop-shaped decorated metal pieces, irregular reddish-brown fragments, a dark oval pendant-like object, polygonal textured fragments, and a long narrow decorated metal object. The paired views reveal substantial differences in texture, ornamentation, boundary shape, surface condition, and thickness among the test objects.
Figure 5. Visual comparison of reconstruction results. (a) Result produced by the proposed method. (b) Reconstruction without thickness constraints, resulting in surface interpenetration. (c) Reconstruction without hypothesis flipping, leading to inaccurate stitching. Three reconstruction results of the same decorated artifact are shown from left to right. The full proposed method produces a continuous surface with consistent alignment around the highlighted region. Removing the thickness constraint causes the reconstructed front and back surfaces to intersect. Removing rotation-hypothesis evaluation produces an incorrect relative orientation and mismatched local structure. Red boxes and enlarged insets emphasize the differences near the registration boundary.
Object
Baseline methods
PSNR ↑
SSIM ↑
LPIPS ↓
IoU →1
AreaRatio →1
Artifact1
FPFH+ICP
12.812
0.477
0.620
0.292
3.134
CPD+ICP
10.668
0.358
0.611
0.303
0.607
FGR+ICP
10.585
0.298
0.545
0.497
1.941
GICP
11.648
0.434
0.564
0.305
3.338
GeoTransformer
12.694
0.436
0.494
0.486
2.379
Predator
13.137
0.483
0.485
0.459
2.166
Table 1. Comparison across different artifact cases. Within each artifact group, the top three results are emphasized by red-colored text and orange-shaded backgrounds with decreasing intensity, where darker shades denote higher-ranked performance.
Figure 6. Sensitivity of mask-guided point cloud purification to different semantic consistency thresholds τ . As τ increases from 0.4 to 0.9, the purified point clouds remain visually stable, with only minor changes near the boundary regions. Six renderings of the same elongated artifact after mask-guided point-cloud purification, corresponding to semantic-consistency thresholds of 0.4, 0.5, 0.6, 0.7, 0.8, and 0.9. The main artifact surface and overall boundary remain nearly unchanged across all thresholds. Increasing the threshold progressively removes a small number of isolated and weakly supported points near the outer boundary while preserving the central structure.
Object
Percentile interval
Front extent
Back extent
Selected hypothesis
Artifact2
1%–99%
0.30191
0.24756
Identity hypothesis
2%–98%
0.27836
0.23128
Identity hypothesis
3%–97%
0.26240
0.22051
Identity hypothesis
5%–95%
0.24273
0.20518
Identity hypothesis
10%–90%
0.20628
0.17403
Identity hypothesis
Artifact4
1%–99%
0.06878
0.13458
Flip-Y hypothesis
Table 2. Effect of percentile intervals on thickness-axis extent estimation and rotation hypothesis selection.
Object
γ
Identity
Flip-X
Flip-Y
Flip-Z
Selected hypothesis
Artifact4
0.5
0.76839
0.84200
0.89447
0.81110
Flip-Y hypothesis
1.0
0.91305
0.92541
0.97597
0.95562
Flip-Y hypothesis
1.5
0.94665
0.94321
0.98011
0.97681
Flip-Y hypothesis
2.0
0.95818
0.95464
0.99229
0.99823
Flip-Z hypothesis
2.5
0.97021
0.96180
0.99949
1.00000
Flip-Z hypothesis
3.0
0.98178
0.97244
1.00000
1.00000
Ambiguous
Table 3. Sensitivity of fitness evaluation to different γ values on representative objects.
Figure 7. Robustness evaluation under different noise perturbations. Two types of noise are introduced to the purified front–back point clouds, and the registration results are compared under increasing noise levels (0%, 1% and 2%). The proposed method successfully aligns and stitches the point clouds even in the presence of noise, demonstrating stable registration performance. Two sequences illustrate registration robustness under increasing noise. The upper sequence adds Gaussian noise at zero, one, and two percent, producing an increasingly diffuse point distribution around the reconstructed surface. The lower sequence adds outlier noise at the same levels, shown as increasingly dense red points surrounding the artifact. In both sequences, the registered surfaces retain their overall alignment and recognizable shape as the noise level increases.
Figure 8. Ablation study on two objects. From left to right: full model, without thickness prior, and without rotation hypothesis evaluation. Removing the thickness constraint leads to noticeable surface interpenetration, while further removing the rotation hypothesis causes severe misalignment between the reconstructed front and back surfaces. Ablation results for two artifacts are arranged in two rows and three columns. Each row progresses from the full model on the left, to the result without the thickness prior in the center, and to the result without rotation-hypothesis evaluation on the right. Removing the thickness prior allows the two surfaces to intersect or collapse toward one another, while additionally removing hypothesis evaluation produces a visibly incorrect orientation and severe front-to-back misalignment.
Global point-cloud registration remains challenging when limited overlap, repetitive geometry, and sensor noise produce correspondence sets dominated by outliers. Planar regions are particularly difficult for conventional point descriptors and are therefore often suppressed or discarded before matching. We present PARTE (Plane-Assisted Robust Transformation Estimation), a global registration method that instead treats planar structure as complementary registration evidence. PARTE extracts planar patches and represents them using our novel Plane Context Histogram (PCH), a descriptor that encodes the geometry surrounding each patch, while a two-level matching procedure identifies reliable plane correspondences. Candidate point and plane correspondences are combined in a confidence-weighted compatibility graph for joint outlier rejection, followed by rigid transformation estimation. When no usable plane correspondences are available, PARTE naturally reduces to point-only registration. We evaluate PARTE on 8,097 registration pairs across six indoor and outdoor benchmarks spanning dense RGB-D and sparse LiDAR measurements. Evaluations show PARTE achieves the highest overall success rate against 13 standard and state-of-the-art methods while maintaining low runtime. An open-source C++ implementation with Python bindings is provided at https://ariarobotics.github.io/parte/.
Department of Computer Science, Colorado School of Mines. · Department of Mechanical Engineering, University of Texas at Dallas. · U.S. Army Combat Capabilities Development Command, Army Research Laboratory
We address the challenge of point cloud registration using color information, where traditional methods relying solely on geometric features often struggle in low-overlap and incomplete scenarios. To overcome these limitations, we propose GeGS-PCR, a novel two-stage method that combines geometric, color, and Gaussian information for robust registration. Our approach incorporates a dedicated color encoder that enhances color features by extracting multi-level geometric and color data from the original point cloud. We introduce the \textbf{Ge}ometric-3D\textbf{GS} module, which encodes the local neighborhood information of colored superpoints to ensure a globally invariant geometric-color context. Leveraging LORA optimization, we maintain high performance while preserving the expressiveness of 3DGS. Additionally, fast differentiable rendering is utilized to refine the registration process, leading to improved convergence. To further enhance performance, we propose a joint photometric loss that exploits both geometric and color features. This enables strong performance in challenging conditions with extremely low point cloud overlap. We validate our method by colorizing the Kitti dataset as ColorKitti and testing on both Color3DMatch and Color3DLoMatch datasets. Our method achieves state-of-the-art performance with \textit{Registration Recall} at 99.9%, \textit{Relative Rotation Error} as low as 0.013, and \textit{Relative Translation Error} as low as 0.024, improving precision by at least a factor of 2.
Jiayi Tian, Haiduo Huang, Tian Xia +2
National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center of Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an, Shaanxi, China
Partial-to-full registration remains challenging due to varying overlap ratios, fluctuating point densities, and the presence of noise. While transformers have shown strong potential for point cloud processing, prior methods typically confine them to global context aggregation, overlooking fine-grained local geometry crucial for accurate correspondence. We propose \emph{GAPR-Net}, a learning-based point cloud registration framework with a coarse-to-fine architecture that combines convolution and transformer modules, in which local and global information is fused between the partial and full point clouds using a cross-attention mechanism. To achieve this, a transformation-invariant point-wise geometric feature representation is proposed, which can robustly capture relative geometric features for individual points with respect to their neighboring points. To evaluate the effectiveness of the proposed approach, experiments are conducted on four geometrically distinct bones, including the tibia, femur, pelvis, and thoracic cartilage. The overall registration recall reaches 94.2%, the method results in a low RMSE of 1.992 mm and R2 values of 0.908 and 0.974 for rotation and translation, respectively. The results demonstrate that the proposed method effectively addresses the partial-to-full point cloud registration problem. The proposed method enables highly accurate 3D point cloud registration using partial observation, providing a critical foundation for precise surgical navigation and robotic interventions in computer-assisted surgery. The code will be accessed after the double-blind review process.
Siyu Zhou, Zhongliang Jiang
The Chair for Computer Aided Medical Procedures, Technical University of Munich, Munich, Germany. · The University of Hong Kong, Hong Kong SAR, China