Source-linked AI summary
Going Further with Point Pair Features
Stefan Hinterstoisser, Vincent Lepetit, Naresh Rajkumar, Kurt Konolige
TL;DR
Drost-PPF is useful for 3D object detection but is inefficient and sensitive to sensor noise and background clutter. This paper introduces improved feature sampling and voting-related processing, with experiments showing state-of-the-art competitiveness and low runtime in challenging settings. Its practical scope is bounded by cases where object pose is unknown and no single smaller sampling volume can preserve all valid point pairs.
Problem
Drost-PPF is inefficient and sensitive to sensor noise and background clutter, which can generate spurious votes and obscure correct hypotheses.
Method
The paper introduces feature-sampling strategies and processing modifications that improve Drost-PPF’s handling of point-pair matching, noise, and clutter.
Results
The improved method outperforms state-of-the-art approaches on object instance detection and pose estimation, including methods using color cues, while operating at low computational cost.
Takeaways & Limitations
Clever feature sampling and feature spreading make PPF-based detection competitive with state-of-the-art methods in challenging cluttered and noisy settings.
Takeaways & Limitations
Without prior object pose information, no single sampling volume smaller than the conservative object-sized volume can preserve all valid point pairs under every configuration.
Abstract
from arXiv · showhide
Point Pair Features is a widely used method to detect 3D objects in point clouds, however they are prone to fail in presence of sensor noise and background clutter. We introduce novel sampling and voting schemes that significantly reduces the influence of clutter and sensor noise. Our experiments show that with our improvements, PPFs become competitive against state-of-the-art methods as it outperforms them on several objects from challenging benchmarks, at a low computational cost.
1 Introduction
Object instance recognition and 3D pose estimation remain challenging in general environments, where robotics applications require efficiency, reliability, and accuracy. The paper improves Drost’s point-pair approach through better sampling and preprocessing/postprocessing modifications, making it competitive with state-of-the-art methods at low computational cost.
- Robotic grasping and manipulation require efficient, reliable, and accurate object instance recognition and 3D pose estimation in general environments.
- Drost’s point-pair voting approach overlooks pair sampling, making it inefficient and often requiring several seconds to run.
- Sensor noise disrupts quantization, while nearby background clutter casts spurious votes that can mask correct votes.
- The paper introduces a more efficient sampling strategy with small pre- and post-processing modifications to reduce these weaknesses.
- The improved approach beats state-of-the-art methods on several objects from challenging datasets while maintaining low computational cost.
- Figure 1 illustrates simultaneous detection under different poses, clutter, partial occlusion, and illumination changes.
2 Related Work
The related work spans sparse feature, template-based, local patch-based, and point-cloud-based methods for 3D object detection and pose estimation. The paper builds on Drost’s point-cloud method, which is suitable for detection but shares computational and clutter-related challenges with other point-cloud approaches.
- Sparse Feature-Based Methods: Sparse feature-based methods are less popular in practical robotics because many objects lack stable feature points due to limited texture.
- Template-based methods: Template-based methods match viewpoint-specific object appearances and can refine 3D pose estimates with ICP, including for textureless objects.
- Local patch-based methods: Local patch-based methods use forest-based voting schemes to detect objects and estimate their 3D poses from image or depth patches.
- Point-cloud-based methods: Point-cloud methods include ICP and feature representations such as point pairs, spin-images, and point-pair histograms.
- Point-cloud-based methods: Point-cloud approaches are usually computationally expensive and have difficulty in scenes with heavy clutter.
- Point-cloud-based methods: Drost’s point-cloud method is the starting point for the paper’s own method.
3 “Drost-PPF” [9]
Drost-PPF detects object poses by matching rigid-motion-invariant point-pair features from models and scenes, then voting in a Hough space. Its pairwise matching is extended by related variants, but remains inefficient and sensitive to clutter and sensor noise, motivating the paper’s sampling improvements.
- Drost-PPF couples Point Pair Features with voting to establish scene-to-model correspondences and estimate object pose and location simultaneously.
- The method computes a four-component PPF from two 3D points and their normals using distance and three inter-vector angles.
- Because PPFs are invariant to rigid motions, the method can detect objects under translations and rotations.
- Discretized PPFs index a lookup table storing model points and rotation angles for use during voting.
- At runtime, scene-point pairs are matched through the lookup table, whose entries vote for model points and rotation angles in a Hough transform.
- Extensions add color, visibility context, planar boundaries, or image gradients, while other methods target symmetry-based speedups and matching performance.
- Drost-PPF has been criticized for high search-space dimensionality, inefficiency, sensitivity to clutter and sensor noise, and weaker dataset performance.
- The paper proposes feature-sampling strategies intended to outperform state-of-the-art methods on standard and challenging occlusion datasets and to speed up matching.
4 Method
The method improves PPF robustness and efficiency through discriminative point-pair sampling, two-stage voting, duplicate-vote suppression, and clustering that limits clutter and repetitive-structure bias.
- Sampling: The method retains nearby point pairs when their normals differ by more than 30 degrees, preserving likely discriminative features.This supplements distance-based subsampling, which otherwise removes close pairs with similar normals.
- Sampling: The method pairs each scene point only with points inside an object-sized voting region instead of all scene points.This avoids unnecessary pairs while retaining pairs that can belong to the same object.
- Sampling: Two voting balls first process a small region based on the object bounding box, then a larger conservative region, reducing cluttered votes without discarding relevant pairs.Peaks from the small ball are extracted before adding pairs accepted only by the large ball.
- Sampling: O(nk) complexity replaces Drost-PPF’s quadratic complexity while guaranteeing that all relevant point pairs are considered.Here, k is usually at least one magnitude smaller than n.
- Voting: A compact bit array indexed by quantized PPFs and scene rotations suppresses repeated votes caused by discretization and spreading.Each 32-bit entry records whether a quantized rotation has already voted for a quantized PPF.
- Voting: The duplicate-vote solution scales linearly with possible quantized PPFs and avoids the direct method’s typical threefold slowdown at 1000 model points.The implementation uses 22 angle bins and 40 distance bins; the direct method becomes significantly slower beyond 650 model points.
- Postprocessing: Bottom-up clustering allows similar pose hypotheses to join multiple clusters while preventing repeated model-point hypotheses from biasing a cluster.This addresses unreliable accumulator vote counts under noisy sensors and background clutter and limits bias from repetitive planar structures.
5 Experimental Results
Experiments on standard and occlusion benchmarks show that the proposed method achieves strong recognition performance using depth alone, while substantially reducing runtime relative to Drost-PPF. Its practical complexity is improved by spatial lookup, although the worst case remains quadratic.
- 5.1 ACCV dataset of [12]: The method performs best on eight of thirteen objects in the ACCV benchmark while using only depth data.The compared methods additionally use color data.
- 5.2 Occlusion Dataset of [16]: On the occlusion dataset, the approach performs better for five of eight objects and averages 3.3% higher recognition than Krull et al.Krull et al. use color data and a ground plane during training, whereas this method uses only depth.
- 5.3 Runtime: Processing a 640×480 depth map takes 0.1s–0.8s, with smaller objects requiring denser subsampling and more processing time.The runtime varies with object diameter because preprocessing samples smaller objects more densely.
- 5.3 Runtime: The approach is about 6 times faster than Drost-PPF while being significantly more accurate.The authors also note that GPU implementation could provide further acceleration.
- 5.4 Worst Case Runtime Discussion: In practice, spatial lookup changes runtime from Drost-PPF’s O(n^2) behavior to O(kn) when k is much smaller than n.The number of scene points can exceed 15k, while k is often over twenty times smaller than n.
- 5.5 Contribution of each Step: Sensor-noise handling contributes 43.1% of the measured gain, followed by smart point sampling at 41.3% and preprocessing at 15.6%.These contributions were evaluated individually against the original Drost-PPF implementation on the occlusion dataset.
6 Conclusion
Clever feature sampling and feature spreading for sensor noise improve Drost’s method, enabling it to outperform state-of-the-art approaches in object detection and pose estimation, including color-based methods.
- Feature sampling and feature spreading for sensor noise boost Drost’s method beyond state-of-the-art approaches in object instance detection and pose estimation.