PaperlayerView exact revision on arXiv

Source-linked AI summary

Geometry-Constrained Bidirectional Point Cloud Registration for Thin, Sheet-Like Heritage Artifacts

Yuezhe Zhang, Lei Wei, Jingnan Du, Shuai Wan

arXiv:2610.07793v1cs.CV

TL;DR

Thin heritage artifacts require separate non-contact reconstruction of their front and back, leaving limited shared geometry for registration. The paper combines thickness-aware constraints with rotation-hypothesis evaluation and reports competitive or improved performance in most tested cases, especially for projected-area consistency and physically plausible alignment.

  • Problem

    Separately reconstructed front and back surfaces have limited shared geometry, while smooth, low-curvature surfaces can make local feature descriptors less distinctive and correspondences unreliable.

    Source: §1 · p. 2, §2.2 · p. 4

  • Method

    The method combines semantic-guided purification and PCA normalization with a physical-thickness prior and rotation hypotheses refined by point-to-plane ICP.

    Source: §1 · p. 2, §5 · p. 24

  • Results

    The method achieves competitive or improved performance in most tested cases, particularly in projected area consistency and physically plausible front–back alignment, though some baselines perform better on image-space or silhouette metrics for some artifacts.

    Source: §5 · p. 24

  • Takeaways & Limitations

    Ablations indicate that removing the thickness prior permits surface interpenetration, while removing rotation-hypothesis evaluation can produce flipped or misoriented alignments.

    Source: §4.4 · p. 23

    Evaluation focuses on consistency with original image observations rather than an additional independent geometric reference model.

    Source: §4.2 · p. 16

Abstract

from arXiv · show

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/.

1 Introduction

Thin heritage artifacts must be reconstructed from independently acquired front and back surfaces, making registration difficult when shared geometry is scarce. The paper proposes a non-contact pipeline that combines semantic purification, PCA normalization, thickness constraints, and rotation-hypothesis search.

1 Introduction

  • Heritage digitization uses active scanning and image-based reconstruction, but active systems can require specialized hardware, controlled settings, and complex calibration.SfM–MVS and close-range photogrammetry are also used, while fragile artifacts may require multi-orientation non-contact acquisition and point-cloud registration.Source: §1 · pp. 1–2
  • Independent front- and back-side acquisition leaves thin artifacts in separate coordinate systems, while near-planar geometry and limited thickness variation make registration ambiguous and unstable.Fragility prevents suspension or artificial markers, so both sides are captured separately under non-contact conditions.Source: §1 · p. 2
  • The method combines semantic-guided preprocessing, PCA-based geometric normalization, and thickness-aware registration for thin, sheet-like heritage artifacts.The proposed pipeline is presented as a stable, physically consistent approach to non-contact front-back reconstruction.Source: §1 · pp. 2–3
  • A percentile-based thickness estimate in PCA-normalized coordinates acts as a geometric prior during registration to prevent non-physical structural collapse.Source: §1 · p. 2
  • A finite set of rotations derived from principal-axis sign permutations is refined with point-to-plane ICP and ranked using a geometry-consistent fitness criterion.Source: §1 · p. 2
  • Image-level semantic masking and 3D spatial clustering purify point clouds before registration, removing background without physical markers.Source: §1 · p. 2

2 Related Work

Prior registration methods span iterative, global, feature-based, and learning-based approaches, but thin and low-overlap structures violate common assumptions about distinctive geometry and shared surface regions. Their near-planarity, weak local features, and physical thickness constraints motivate specialized registration strategies.

2.1 Point Cloud Registration

  • Registration research includes local iterative methods such as ICP, global methods such as 4PCS, and later approaches including Super4PCS.The broader field also includes handcrafted feature-based and learning-based methods.Source: §2.1 · p. 3
  • Feature-based registration uses local descriptors such as FPFH, USC, and SHOT to establish point-cloud correspondences.Source: §2.1 · p. 3
  • Learning-based methods include PointNetLK, FCGF, D3Feat, Predator, and GeoTransformer, which learn representations or improve correspondence estimation.Source: §2.1 · p. 4

2.2 Registration of Thin and Low-Overlap Structures

  • Many registration methods assume rich 3D geometry and overlapping surfaces, conditions often absent in thin, sheet-like artifacts and near-planar structures.Source: §2.2 · p. 4, §2.1 · p. 4
  • Near-planar point clouds have anisotropic covariance and can yield rank-deficient ICP optimization, causing unstable convergence or ambiguous transformations.Planar or symmetric structures also reduce rigid-transformation observability.Source: §2.2 · p. 4
  • Smooth, low-curvature surfaces weaken the distinctiveness of local descriptors such as FPFH and SHOT, making correspondences unreliable.Source: §2.2 · p. 4
  • Front and back surfaces may share little or no overlap, while their small physical separation means unconstrained registration can cause interpenetration or thickness-direction collapse.Thickness constraints and canonical normalization are identified as priors that can stabilize under-constrained alignment.Source: §2.2 · pp. 4–5

3 Proposed Method

The proposed method purifies the front- and back-side point clouds, uses PCA-based normalization to estimate artifact thickness, and evaluates orientation hypotheses refined with thickness-initialized point-to-plane ICP. It selects the hypothesis with the highest thickness-aware fitness score and maps its transformation back to the original coordinates to produce a stitched artifact model.

3.1 Semantic-Guided 2D–3D Purification

  • Independent front- and back-side purification combines multi-view semantic filtering with geometric cleanup to produce structurally consistent artifact-only clouds without cross-side correspondences.Statistical outlier removal and dominant-component retention address residual points and background fragments that semantic masks may leave behind, including under near-zero overlap.Source: §3.1 · pp. 5, 8, 9

3.2 PCA-Based Canonical Normalization and Thickness Estimation

  • PCA canonicalization aligns each cloud to intrinsic axes, using the least-variance axis as a thickness direction and robust statistics to estimate scale and thickness.A right-handed basis corrects PCA sign ambiguity; percentile-based thickness bounds reduce sensitivity to reconstruction noise, and the resulting half-thickness offset serves as a geometric prior without cross-side correspondences.Source: §3.2 · pp. 10–11

3.3 Geometry-Constrained Bidirectional Registration

  • The registration stage tests finite global orientation hypotheses, initializes each with a thickness-derived offset, and refines each candidate using point-to-plane ICP.For double-sided acquisition, the plausible orientations are restricted to 180° rotations around canonical axes; the pipeline selects the optimal transformation with a thickness-aware fitness criterion.Source: Figure 1 · p. 6, §3.3 · p. 12
  • The implementation sets γ to 2.0 as a balance between rejecting valid noisy correspondences with a stricter threshold and losing orientation discrimination with a looser one.The sensitivity of γ is analyzed in Section 4.3.3.Source: §3.3 · p. 13
  • The highest-scoring thickness-aware fitness hypothesis is mapped from PCA space to the original coordinates, registering the back cloud with the front surface to produce the stitched model.Source: §3.3 · pp. 12–14

4 Experimental Results

The experiments compare front–back registration methods using image- and projection-based evidence, then examine parameter sensitivity, noise robustness, and component ablations. They show that the full method’s thickness constraint and rotation-hypothesis evaluation support more physically plausible and stable stitching.

4.1 Experimental Setup

  • Front and back surfaces were acquired independently from multiple viewpoints, and all methods used the same purified front–back point-cloud pairs.The artifacts rested on a flat supporting surface during non-contact acquisition.Source: §4.1 · p. 14

4.2 Comparison with Existing Registration Methods

  • Registered clouds were rendered from calibrated reconstruction viewpoints and assessed against captured images and object masks, not an independent 3D reference model.The authors leave more extensive reference-based 3D geometric validation for future work.Source: §4.2 · p. 16
  • Image-space and projection metrics provide complementary evidence, while outcomes vary with artifact geometry; Fig. 5 contrasts the proposed and degraded reconstructions.IoU measures silhouette overlap, whereas AreaRatio reflects projected area consistency.Source: §4.2 · p. 17, Figure 5 · p. 17

4.3 Parameter Sensitivity and Robustness Evaluation

  • Across τ values from 0.4 to 0.9, the main artifact structure remained stable, with changes chiefly near boundaries; τ=1.0 was treated as degenerate.At τ=1.0, purification would retain only a very small number of points.Source: §4.3.1 · p. 19, Figure 6 · p. 19
  • The 2%–98% percentile interval balanced removal of sparse extremes against excessive trimming for thickness-axis extent estimation.Narrower intervals generally decrease the estimated front- and back-side extents; wide intervals may retain noise or boundary artifacts.Source: §4.3.2 · p. 20
  • Smaller γ values lowered fitness scores, while excessive relaxation caused hypothesis ambiguity; γ=2.0 was adopted as the empirical default.Objects 4 and 6 became ambiguous at γ=3.0, and Object 6 also at γ=2.5.Source: §4.3.3 · pp. 21–22
  • The method aligned and stitched point clouds at synthetic noise levels of 0%, 1%, and 2% of scene scale.The reported alignment remained stable without structural misalignment or failure across these tested levels.Source: §4.3.4 · p. 22, Figure 7 · p. 22

4.4 Ablation Study

  • Ablations showed that removing thickness constraints led to surface interpenetration, while also removing rotation-hypothesis evaluation caused severe front–back misalignment.The full method preserved more plausible thickness separation and visually more stable stitching.Source: §4.4 · pp. 23–24, Figure 8 · p. 23

5 Conclusion and Future Work

The paper presents a geometry-constrained registration pipeline for independently reconstructed thin-sheet surfaces, combining semantic purification, PCA normalization, and thickness-aware alignment. It reports mostly competitive or improved results, while noting unresolved ambiguity in strongly symmetric or weak-boundary artifacts and proposing direct 3D validation and broader extensions.

5 Conclusion and Future Work

  • The pipeline combines semantic-guided purification, PCA-based canonical normalization, thickness-aware registration, rotation hypothesis evaluation, and point-to-plane ICP refinement.Source: §5 · p. 24
  • Results were competitive or improved in most tested cases, especially in projected-area consistency and physically plausible front–back alignment.Some baselines still score better on image-space or silhouette metrics for particular artifacts.Source: §5 · p. 24
  • Strong geometric symmetry or weak boundary cues can leave front–back orientation ambiguous and produce inaccurate stitching.Source: §5 · p. 24
  • Future work proposes known-geometry replicas for direct 3D validation, neural rendering integration, and multi-fragment alignment.Source: §5 · p. 24

Related papers

Loading 2610.07793v1…