Abstract
This paper introduces a novel, watermark-free approach for detecting plagiarized or copied 3D point clouds. By aligning two point clouds and analyzing the probability matrix of their point correspondences using techniques like Robust PCA, the method accurately identifies manipulated copies—including those subjected to noise, smoothing, or rotation—without relying on fragile embedded signals. Experimental results on the HM25 dataset demonstrate that this intrinsic statistical comparison significantly outperforms traditional watermarking and deep-learning-based shape retrieval methods in copy detection tasks.
Introduction
Protecting the copyright of high-quality 3D models, such as those used in gaming and cultural heritage, is a significant challenge. The standard defense, 3D watermarking, embeds a hidden signal into the geometry. However, these signals are often destroyed by common operations like noise addition, cropping, or remeshing. Alternatively, 3D shape retrieval methods are designed to find objects of the same category rather than identifying specific, manipulated copies. This research bridges this gap by treating copy detection as a direct similarity measurement problem robust to aggressive geometric and topological attacks.
Methods
The proposed detection pipeline consists of several stages designed to handle dense point clouds efficiently. The process begins with Hierarchical EM (HEM) downsampling to preserve geometric features while reducing computational load, followed by rigid alignment.
The core methodology involves:
- Point Set Registration: The authors use Coherent Point Drift (CPD), a Gaussian Mixture Model (GMM)-based algorithm, to rigidly align the suspect model with the original.
- Similarity Distance Calculation: During alignment, a probability matrix representing point correspondences is generated. Three novel metrics are used to extract similarity:
- Low-Rank (LR) Measure: Uses Robust PCA (RPCA) to isolate structural relationships from noise.
- Kurtosis (KURT) Measure: Analyzes the statistical sharpness of correspondence peaks.
- Correlation (CORR) Measure: Calculates Pearson distance based on reconstructed geometric correspondences.
To scale these operations, a segmentation-and-fusion approach splits the aligned models into chunks for parallel processing.
Key Insights
A fundamental insight of this work is that the internal probability matrix of a GMM-based registration contains vital structural data. When points are properly aligned, a true copy manifests as a distinct diagonal band in this matrix. By applying RPCA—a technique typically used in image denoising—the authors can effectively strip away the "noise" of geometric manipulations to isolate the underlying structural identity of the model.
Results and Discussion
The method was evaluated against traditional watermarking and modern shape retrieval techniques.
- Robustness to Attacks: The Correlation (CORR) metric significantly outperformed Euclidean distance-based watermarking. For instance, under a 0.5% noise attack, the watermark's Bit Error Rate (BER) failed at 0.31, while the proposed method maintained perfect detection (1.0).
- Retrieval Performance: On the newly introduced HM25 dataset—containing 13,750 models with various geometric attacks—the method achieved a 60.8% Top-1 retrieval rate. This far exceeds deep learning benchmarks like MVCNN (28.8%) and MeshNet (26.4%), which often confuse similar-looking objects with actual copies.
Conclusion
This research demonstrates that intrinsic statistical comparison of point clouds is a highly effective, watermark-free alternative for copy detection. While the computational cost of SVD and CPD remains a challenge, the use of hierarchical downsampling and parallel segmentation provides a path toward scalable implementation. Future work will focus on improving resilience to severe cropping and optimizing processing speeds for massive databases.
Summary updated