Rotational Projection Statistics for 3D Local Surface Description and Object Recognition

Yulan Guo, Ferdous Ahmed Sohel, Mohammed Bennamoun, Min Lu, Jianwei Wan

Abstract

Recognizing 3D objects in the presence of noise, varying mesh resolution, occlusion and clutter is a very challenging task. This paper presents a novel method named Rotational Projection Statistics (RoPS). It has three major modules: Local Reference Frame (LRF) definition, RoPS feature description and 3D object recognition. We propose a novel technique to define the LRF by calculating the scatter matrix of all points lying on the local surface. RoPS feature descriptors are obtained by rotationally projecting the neighboring points of a feature point onto 2D planes and calculating a set of statistics (including low-order central moments and entropy) of the distribution of these projected points. Using the proposed LRF and RoPS descriptor, we present a hierarchical 3D object recognition algorithm. The performance of the proposed LRF, RoPS descriptor and object recognition algorithm was rigorously tested on a number of popular and publicly available datasets. Our proposed techniques exhibited superior performance compared to existing techniques. We also showed that our method is robust with respect to noise and varying mesh resolution. Our RoPS based algorithm achieved recognition rates of 100%, 98.9%, 95.4% and 96.0% respectively when tested on the Bologna, UWA, Queen's and Ca' Foscari Venezia Datasets.

Introduction

3D object recognition is a fundamental challenge in computer vision, particularly when dealing with real-world sensor data. Scanned point clouds and meshes often suffer from significant noise, varying point densities (mesh resolution), and partial visibility due to clutter and occlusion. Traditional descriptors like Spin Images or SHOT frequently struggle under these conditions, often due to an unstable Local Reference Frame (LRF) or reliance on sensitive geometric features like surface normals. This paper introduces Rotational Projection Statistics (RoPS), a robust local surface descriptor designed to overcome these limitations by providing high discriminative power alongside extreme resilience to data imperfections.

Methodology

Local Reference Frame (LRF) Definition

A critical step in local surface description is establishing a consistent coordinate system. Unlike previous methods that rely solely on mesh vertices, the proposed LRF calculation utilizes the continuous surface of the triangles within a local patch. By calculating a scatter matrix based on all points lying on these local surface triangles, the method achieves a much higher degree of repeatability.

The LRF is calculated using the continuous surface of the triangles, making it more stable against resolution changes.

The LRF is derived through eigenvalue decomposition of this scatter matrix, followed by a unique sign disambiguation technique to ensure a consistent orientation across different scans.

RoPS Feature Descriptor

Once the LRF is established, the local surface patch is processed to generate the RoPS descriptor. This involves:

  1. Rotation: The patch is rotated around the x, y, and z axes of the LRF to capture multiple viewpoints.
  2. Projection: For each rotation, points are projected onto the 2D planes (xy, xz, yz).
  3. Statistical Summarization: Instead of simple histograms, the distribution of projected points is summarized using low-order central moments and Shannon entropy. This statistical approach compresses complex 3D geometry into a compact, noise-resistant vector.

Experimental Evaluation

The RoPS algorithm was evaluated across several challenging public datasets, including the Bologna, UWA, Queen’s, and Ca’ Foscari Venezia datasets. These datasets provide a rigorous testing ground for robustness against noise, decimation, and occlusion.

Sample 3D models used in the evaluation, representing a variety of geometric complexities.

Key Results

The proposed method demonstrated state-of-the-art performance:

  • High Accuracy: Recognition rates reached 100% on the Bologna dataset and remained above 95% for the others.
  • Robustness to Decimation: RoPS maintained over 90% accuracy even when mesh density was reduced to 1/8th, significantly outperforming competitors.
  • Noise Resilience: The descriptor remained effective under Gaussian noise levels that caused other methods to fail.
  • Efficient Pose Estimation: The stability of the LRF allows for 3D pose estimation using only a single correct feature match, avoiding the computational cost of multi-match verification.

Conclusion

Rotational Projection Statistics (RoPS) provides a powerful framework for 3D object recognition. By leveraging the continuous surface for LRF definition and statistical projections for feature description, it achieves remarkable robustness to common data artifacts. Future work includes integrating photometric data to enhance discriminability for geometrically similar objects and extending the method to large-scale 3D shape retrieval.

Summary updated

Abstract

Recognizing 3D objects in the presence of noise, varying mesh resolution, occlusion and clutter is a very challenging task. This paper presents a novel method named Rotational Projection Statistics (RoPS). It has three major modules: Local Reference Frame (LRF) definition, RoPS feature description and 3D object recognition. We propose a novel technique to define the LRF by calculating the scatter matrix of all points lying on the local surface. RoPS feature descriptors are obtained by rotationally projecting the neighboring points of a feature point onto 2D planes and calculating a set of statistics (including low-order central moments and entropy) of the distribution of these projected points. Using the proposed LRF and RoPS descriptor, we present a hierarchical 3D object recognition algorithm. The performance of the proposed LRF, RoPS descriptor and object recognition algorithm was rigorously tested on a number of popular and publicly available datasets. Our proposed techniques exhibited superior performance compared to existing techniques. We also showed that our method is robust with respect to noise and varying mesh resolution. Our RoPS based algorithm achieved recognition rates of 100%, 98.9%, 95.4% and 96.0% respectively when tested on the Bologna, UWA, Queen's and Ca' Foscari Venezia Datasets.
BibTeX
@article{guo2013rotational,
  author = {Yulan Guo and Ferdous Ahmed Sohel and Mohammed Bennamoun and Min Lu and Jianwei Wan},
  title = {{Rotational Projection Statistics for 3D Local Surface Description and Object Recognition}},
  journal = {Int. J. Comput. Vis.},
  year = {2013},
  url = {https://openreview.net/forum?id=UU5PoDI0WDW},
}
Figure