Source-linked AI summary
Rotational Projection Statistics for 3D Local Surface Description and Object Recognition
Yulan Guo, Ferdous Sohel, Mohammed Bennamoun, Min Lu, Jianwei Wan
TL;DR
3D object recognition is difficult under noise, changing mesh resolution, occlusion, and clutter. The paper introduces RoPS with an all-point LRF and hierarchical recognizer, reporting strong robustness and superior results across four datasets.
Problem
3D object recognition remains challenging under noise, varying mesh resolution, occlusion, and clutter, while existing descriptors can have limited descriptiveness or robustness.
Method
The method combines an LRF built from all local-surface points, RoPS rotational projection statistics, and a hierarchical 3D object recognition algorithm.
Results
The proposed techniques show high descriptiveness and robustness to noise, varying mesh resolution, holes, and other deformations, with superior recognition results across four datasets.
Takeaways & Limitations
RoPS provides a local surface description and recognition approach evaluated as robust across several nuisances and complex-scene recognition settings.
Takeaways & Limitations
Local symmetrical surfaces can produce LRF sign ambiguity, leading the method to avoid such surfaces through eigenvalue-ratio-based feature selection.
Abstract
from arXiv · showhide
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.
1 Introduction
3D object recognition remains difficult under noise, changing mesh resolution, occlusion, and clutter. The paper addresses these challenges with a robust LRF, RoPS descriptor, and hierarchical recognition algorithm.
- 3D object recognition from range images remains challenging under noise, varying mesh resolution, occlusion, and clutter.
- The proposed LRF uses eigenvalue decomposition of the scatter matrix of all local-surface points with sign disambiguation.
- RoPS rotationally projects neighboring points onto three coordinate planes and computes statistics including central moments and entropy.
- The paper presents a hierarchical recognition algorithm and evaluates the proposed techniques on four popular datasets.
- Feature matching is critical because it affects transformation accuracy, verification and refinement time, while descriptors must remain robust to several nuisances.
2 Related Work
Prior local descriptors and recognition algorithms face limitations in spatial encoding, LRF repeatability, robustness, and computational demands. The paper proposes an all-point LRF, RoPS descriptor, and hierarchical recognizer designed to address these gaps.
- Features without LRF: Descriptors without an LRF often histogram geometric attributes, discarding spatial information and limiting descriptiveness.
- Features with LRF: A reference axis is incomplete because rotation around the normal remains uncertain, while existing LRFs may lack uniqueness, robustness, or unambiguous orientation.
- 3D Object Recognition: Existing recognition algorithms commonly require three feature correspondences, increasing runtime through combinatorial matching and reducing transformation precision.
- Contributions: The proposed LRF uses all local-surface points rather than only mesh vertices and adds sign disambiguation for uniqueness and robustness.
- Contributions: RoPS encodes neighboring-point distributions through rotational projections onto three coordinate planes and statistical features.
- Contributions: The hierarchical recognizer requires only one correct feature correspondence and can also work with existing descriptors such as spin images.
3 Local Reference Frame
The paper defines a unique, repeatable local reference frame from all points on a cropped local surface, using weighted scatter matrices, eigenvectors, and sign disambiguation. Its evaluation reports improved LRF robustness and performance, while local symmetry remains a source of large errors addressed through feature-point selection.
- 3.1 Coordinate Axis Construction: A local surface is cropped within radius r, and points inside each triangle are represented from its three vertices.For triangle vertices p_i1, p_i2, and p_i3, the representation uses parameters satisfying 0 ≤ s,t ≤ 1 and s + t ≤ 1.
- 3.1 Coordinate Axis Construction: The overall scatter matrix combines triangle scatter matrices using area and feature-point-distance weights.The area weight compensates for triangle-area variation, while the distance weight reduces contributions from distant points.
- 3.1 Coordinate Axis Construction: Eigenvalue decomposition of the scatter matrix provides three orthogonal axes ordered by decreasing eigenvalue magnitude, but their signs require disambiguation.The proposed method uses all local-surface points rather than only mesh vertices, targeting robustness to noise and varying mesh resolution.
- 3.1 Coordinate Axis Construction: Each eigenvector’s sign is aligned with the majority direction of scatter vectors, and the third axis is obtained by a cross product to form a unique LRF.The feature point is the origin, and the three disambiguated vectors define the x, y, and z axes.
- 3.3 Performance of the Proposed LRF: 83.5% of point pairs were placed in the first LRF-error histogram bin, and the proposed technique performed best among the compared methods.Using both quadratic distance and area weights produced more robust results than using only a linear distance weight; local symmetry caused some errors above 80 degrees.
4 Local Surface Description
RoPS describes local surfaces by rotating LRF-aligned neighborhoods, projecting them onto three planes, and encoding normalized point distributions with statistics. Parameter studies select settings that balance descriptiveness, robustness, and efficiency.
- RoPS Feature Descriptor: RoPS rotates an LRF-aligned local surface, projects neighboring points onto three coordinate planes, and computes statistics from the resulting distributions.The three planes are xy, xz, and yz; statistics from the three distribution matrices form a sub-feature for each rotation.
- RoPS Feature Descriptor: Normalizing each L×L distribution matrix so its bins sum to one provides invariance to varying mesh resolution.The matrix counts projected points in spatial bins before normalization.
- RoPS Feature Descriptor: Central moments and Shannon entropy condense distribution matrices into descriptors that retain geometric information while improving computational and storage efficiency.The selected statistics include low-order central moments and entropy.
- RoPS Feature Descriptor: RoPS concatenates sub-features across rotations, encoding local-surface geometry from multiple viewpoints for increased discrimination.The paper contrasts this local, view-based encoding with global view-based methods that are sensitive to occlusion and clutter.
- Combination of Statistics: The No.6 statistics combination achieved the best tuning performance and was favored because it combined high descriptiveness, noise robustness, and a compact descriptor.Its low-order moments and entropy were considered especially informative and more robust than high-order moments.
- Generation Parameters: Five partition bins offered the selected tradeoff: performance improved from three to five bins but degraded with denser partitions because spatial variation increased sensitivity.The paper therefore used L = 5, while increasing rotations generally improved matching and a support radius of 15mr produced precision and recall of about 0.9.
5 Performance of the RoPS Descriptor
RoPS was evaluated against existing descriptors on benchmark data under noise, mesh decimation, and geometric transformations. It generally provided stronger robustness, including under combined degradations, while remaining compact.
- Bologna Dataset: The Bologna evaluation compared RoPS with five state-of-the-art descriptors using RP Curves on synthetically cluttered scenes with known transformations.The scenes contained six models and 45 scenes, with models randomly rotated and translated to create clutter and pose variation.
- Noise Robustness: Under increasing noise, RoPS achieved the best performance in most cases and was significantly better than SHOT above 0.3mr noise deviation.SHOT was slightly better at 0.2mr and 0.3mr, while RoPS performed better at 0.1mr and at higher noise levels.
- Mesh Resolution: RoPS outperformed all other descriptors across mesh-decimation levels, retaining precision and recall above 0.7 at 1/8 of the original mesh resolution.At 1/8 resolution, its performance was comparable to the best competing results at 1/2 resolution.
- Combined Degradations: With simultaneous noise and mesh decimation, RoPS achieved precision and recall of about 0.9 and significantly outperformed the other methods.NormHist and SHOT followed RoPS, whereas THRIFT failed to work.
- Efficiency: RoPS used a 135-element descriptor, shorter than spin image, NormHist, LSP, and SHOT, making feature matching more efficient than with those descriptors.The cited competing descriptor lengths were 225, 225, 225, and 320, respectively.
- PHOTOMESH Dataset: On PHOTOMESH transformations, RoPS was invariant to color noise, color shot noise, rotation, and scale, while normalized L2 distances stayed at or below 0.06 for several geometric transformations.Sampling was the largest challenge, with average normalized L2 distance increasing from 0.01 to 0.06 across strength levels 1 to 5.
6 3D Object Recognition Algorithm
The algorithm uses the LRF and RoPS descriptor within a hierarchical recognition pipeline that generates, ranks, estimates, verifies, and refines object hypotheses. Model features are represented and indexed offline, while scene processing and recognition operate online.
- The hierarchical algorithm comprises model representation, candidate model generation, transformation hypothesis generation, verification, and segmentation.These modules form the recognition pipeline illustrated in Fig. 10.
- Model features are calculated and indexed with a k-d tree offline, whereas the subsequent recognition modules operate online.Offline processing constructs the model library and accelerates later feature matching.
- Scene features are matched against model features, and vote-ranked models become candidate models for subsequent processing.A correspondence is established when the nearest-to-second-nearest distance ratio falls below threshold τf.
- A single feature correspondence can estimate a rigid transformation by aligning the model and scene local reference frames.This avoids the need for the at-least-three correspondences required by several cited methods.
- Plausible transformations are clustered using Euler-angle proximity and ranked by confidence based on correspondence count and average feature distance.Clusters below half the maximum confidence score are pruned before valid hypotheses are selected.
- Candidate hypotheses are refined with ICP and accepted only when residual error and visible proportion satisfy the flexible thresholding scheme.Accepted model correspondences are removed from the scene, and verification continues across candidates until processing ends or too few scene points remain.
7 Performance of 3D Object Recognition
RoPS-based recognition was evaluated across four datasets containing occlusion and clutter, with additional tests of noise and mesh-resolution variation. The algorithm achieved strong recognition performance and, on reported comparisons, outperformed existing methods.
- Bologna Dataset: 100% recognition rate was achieved by the RoPS-based algorithm under all tested levels of mesh decimation on the Bologna Dataset.NormHist and spin image reached 97.8% and 91.1%, respectively, at 1/8 of the original mesh resolution.
- UWA Dataset: 98.9% overall average recognition rate was achieved on the UWA Dataset, with 100% recognition up to 80% occlusion and 93.1% at 85% occlusion.Only two of 188 objects were not correctly recognized, and no false positives occurred.
- Scale evaluation: Adaptive-scale RoPS performed marginally worse than fixed-scale RoPS because scale-estimation errors adversely affected feature matching and object recognition.The adaptive-scale version nevertheless performed better than the result reported in the cited prior work.
- Queen’s Dataset: 95.4% average recognition rate was achieved on the Queen’s Dataset, compared with 82.4% for the EM-based algorithm.RoPS achieved the best recognition rates for all five models and recognized objects under large amounts of occlusion and clutter.
- Dataset comparison: 98.9% recognition on UWA and 95.4% on Queen’s showed consistent performance across noise, mesh-resolution variation and occlusion.The Queen’s Dataset was more challenging because it was noisier and its points were not uniformly distributed.
- Ca’ Foscari Venezia Dataset: 99% average precision and 96% average recall were achieved on the Ca’ Foscari Venezia Dataset.Average precision exceeded the cited prior result by 6%, while average recall was 1 percentage point higher.
8 Conclusion
The paper introduces RoPS for 3D local surface description and a hierarchical recognition algorithm, supported by a novel local reference frame. Experiments report strong descriptor robustness and improved recognition across challenging datasets, while future work targets photometric cues and shape retrieval.
- RoPS describes local surfaces by rotationally projecting neighboring points onto three coordinate planes and computing distribution statistics.
- The proposed local reference frame uses the scatter matrix of all points on the local surface rather than only mesh vertices.
- Experiments show RoPS has high descriptiveness and strong robustness to noise, varying mesh resolution, holes, and other deformations.
- Recognition experiments report effectiveness under noise and mesh-resolution variation, robustness to occlusion, and improvements over existing algorithms across the tested datasets.
- Future work includes combining geometric and photometric information and applying RoPS to large-scale 3D shape retrieval.