PaperlayerView exact revision on arXiv

Source-linked AI summary

GUARD: Geometric Uncertainty-Aware Point Cloud Denoising and Segmentation for Robotic Hard Disk Drive Disassembly

Zuoxu Wang, Xiao Liang

arXiv:2610.09068v1cs.CVcs.RO

TL;DR

Structured artifacts in HDD point clouds can resemble valid components while receiving plausible semantic labels, so semantic confidence alone does not establish geometric reliability. GUARD combines geometric uncertainty with predictive entropy to filter points and segment parts in a single pass. It improves PointNet++ segmentation on real HDD scans, while the analyses distinguish geometric reliability from semantic confidence and reveal a filtering tradeoff.

  • Problem

    Structured reconstruction artifacts can resemble real components and receive plausible semantic labels, so semantic prediction confidence alone does not establish whether measured geometry is reliable.

    Source: §I · p. 1, D. Hybrid Uncertainty Fusion for… · p. 5

  • Method

    GUARD combines multi-scale geometric representations, Gaussian Process-based geometric uncertainty, and semantic entropy in a single-pass framework for point filtering and segmentation.

    Source: §I · p. 2

  • Results

    GUARD improves PointNet++ part segmentation on real HDD scans, and experiments characterize corrupted-point detection and filtering behavior across benchmarks and backbones.

    Source: D. Ablation Experiment · p. 9, §VI · p. 12

  • Takeaways & Limitations

    Geometric reliability and semantic confidence capture different aspects of point-cloud quality, and artifact suppression must be considered alongside preservation of informative geometry.

    Source: §V · p. 11, §VI · p. 12

    The GP head is less sensitive to minor ghost clusters near real surfaces, which can overlap valid geometry and share similar local patterns.

    Source: D. Hybrid Uncertainty Fusion for… · p. 5

Abstract

from arXiv · show

Reliable robotic disassembly requires part-level representations that distinguish genuine component geometry from scanning and reconstruction artifacts. In point clouds of hard disk drives (HDDs), structured ghost artifacts can resemble valid components locally while remaining inconsistent with the overall geometry, allowing erroneous measurements to receive plausible semantic labels. This creates an engineering information problem: semantic prediction confidence alone does not establish whether the underlying geometry is reliable. We propose \textbf{GUARD}, a geometric uncertainty-aware framework that performs point filtering and segmentation within a single forward pass by modeling the reliability of learned geometric representations. GUARD combines a multi-scale geometric transformer with a multi-bandwidth random Fourier feature Gaussian Process to estimate per-point geometric uncertainty, complemented by predictive entropy to suppress unreliable measurements while preserving informative structures. Evaluation on 2,745 real HDD point clouds shows that GUARD improves PointNet++ segmentation mean intersection over union from 0.7739 to 0.8318. Additional experiments on ShapeNetPart and ScanNet examine robustness across corruption types, point-cloud domains, and segmentation backbones. On manually annotated ScanNet samples, geometric uncertainty achieves a corrupted-point detection F1 score of 0.7931, compared with 0.2212 for predictive entropy. The results demonstrate the value of distinguishing geometric reliability from semantic confidence and reveal a tradeoff between artifact suppression and preservation of informative structures. GUARD contributes a reliability-aware approach to interpreting imperfect 3D measurements for component identification and subsequent robotic handling. Project website: https://001-wang.github.io/GUARD_Point_denoiser/.

I. INTRODUCTION

Robotic HDD disassembly needs accurate part-level geometry, but structured scan artifacts can resemble real components and receive plausible labels. GUARD couples segmentation with geometric reliability assessment to filter unreliable points while preserving useful structure.

I. INTRODUCTION

  • Reliable perception and part-level point-cloud segmentation support robotic identification and manipulation of densely arranged HDD components.HDDs contain reusable or valuable components requiring identification, handling, sorting, and disassembly.Source: §I · p. 1
  • Structured ghost surfaces can resemble valid components locally, yet remain inconsistent with object geometry and receive plausible but incorrect semantic labels.Point-level annotations for these artifacts are expensive and impractical, while preserving valid fine-grained structures remains challenging.Source: §I · p. 1
  • GUARD uses multi-scale geometric representations, geometry-conditioned uncertainty, and semantic prediction entropy to assess measurement reliability alongside segmentation.The framework uses learned geometric feature support to guide filtering in a single pass.Source: §I · p. 2
  • The study evaluates GUARD on real HDD scans and uses ShapeNetPart and ScanNet to examine robustness beyond the HDD setting.The benchmarks include synthetic sensor noise and ghost-cluster corruption, as well as large-scale indoor scenes.Source: §I · p. 2
  • The analysis distinguishes geometric reliability from semantic confidence and examines when artifact removal helps segmentation or removes informative points.The comparisons are intended to characterize filtering tradeoffs and inform reliability-aware filtering strategies.Source: §I · p. 2

II. RELATED WORK

Prior work spans disassembly planning, point-cloud segmentation and denoising, and uncertainty estimation, while GUARD applies reliability modeling to geometric features. Its overview combines geometric and semantic uncertainty, with local and global feature extraction for point-cloud interpretation.

II. RELATED WORK

  • Prior work addresses disassembly planning, segmentation backbones, denoising, and uncertainty, but integration for locally plausible corrupted measurements remains a challenge.Some existing approaches require noise annotations or multiple passes; scalable uncertainty estimation for large-scale 3D segmentation is described as underexplored.Source: §II · pp. 2–3

A. Overview

  • GUARD combines a multi-scale geometric encoder and a Gaussian Process uncertainty head with semantic entropy to assess geometric reliability and semantic ambiguity.The point-wise reliability measure guides filtering while the segmentation backbone predicts part labels.Source: A. Overview · p. 3

B. Geometric Feature Extraction

  • The geometric feature stage uses raw coordinates and applies spectral normalization to stabilize their mapping into latent features.Relative offsets and distances describe neighbors at scales of 16, 32, and 64, while chunked KNN reduces GPU memory use without altering topology.Source: B. Geometric Feature Extraction · p. 3
  • A geometry-aware transformer models interactions among all points to capture global context that can expose locally plausible but globally inconsistent ghost clusters.The architecture includes a multi-scale geometric encoder, geometry-aware transformer, and uncertainty-aware GP head, and can attach to PointNet++ or DGCNN backbones.Source: B. Geometric Feature Extraction · p. 3, Figure 1 · p. 4

C. Multi-bandwidth Geometric Gaussian Process Head (GGPH)

The geometric transformer is deterministic, so GUARD adds a multi-bandwidth random Fourier feature Gaussian Process head to estimate uncertainty in geometric feature space. The resulting predictive variance represents how well a point's geometry is supported by the learned distribution.

C. Multi-bandwidth Geometric Gaussian Process Head (GGPH)

  • The deterministic geometric transformer cannot quantify uncertainty for unseen or noisy structures, motivating a probabilistic head over geometric features.The paper's HDD variance visualization shows most samples in a low-uncertainty region, with deviations corresponding to abnormal or noisy structures.Source: C. Multi-bandwidth Geometric Gaussian… · p. 4, Figure 2 · p. 5
  • MultiRFF-GP approximates a Gaussian Process kernel in geometric feature space using random Fourier features at multiple frequency bands.The three characteristic bandwidths are 0.3, 0.7, and 1.4, intended to represent distinct spatial scales.Source: C. Multi-bandwidth Geometric Gaussian… · p. 4
  • Concatenated multi-bandwidth features jointly represent local shape detail and global geometry for uncertainty modeling.The head operates on spectrally normalized geometric embeddings and models fine-to-coarse geometric uncertainty.Source: C. Multi-bandwidth Geometric Gaussian… · p. 4
  • High predictive variance indicates geometric representations poorly supported by the training distribution, including outliers or ghost clusters.The GP head's feature second-moment estimate is used to form a regularized precision matrix for inference.Source: C. Multi-bandwidth Geometric Gaussian… · p. 4

D. Hybrid Uncertainty Fusion for Robust Segmentation

GUARD combines geometric variance and semantic entropy because distant ghost clusters and near-surface artifacts exhibit different uncertainty patterns. Their weighted fusion is designed to balance structural reliability with semantic ambiguity.

D. Hybrid Uncertainty Fusion for Robust Segmentation

  • Geometric variance identifies major, globally inconsistent ghost clusters but is less sensitive to minor artifacts near valid surfaces.Near-surface clusters may overlap real geometry and share local patterns, and their manual removal can vary across datasets.Source: D. Hybrid Uncertainty Fusion for… · p. 5
  • Predictive entropy captures ambiguous logits near true surfaces, complementing variance for close ghost clusters that geometric uncertainty alone may miss.The paper combines the two signals with weighted coefficients rather than a cascade to balance distinct uncertainty characteristics.Source: D. Hybrid Uncertainty Fusion for… · p. 5
  • The default fusion weight α=0.7 emphasizes geometric uncertainty while retaining an entropy contribution for ambiguous boundary noise.The resulting map is described as supporting downweighting of unreliable points under scanning perturbations.Source: D. Hybrid Uncertainty Fusion for… · p. 5

E. Implementation Details

GUARD integrates the MultiRFF-GP head into PointNet++ and DGCNN, accumulating the precision matrix during training and freezing it for testing.

E. Implementation Details

  • GUARD integrates the MultiRFF-GP head into both PointNet++ and DGCNN backbones.The precision matrix is accumulated online during training and frozen during testing for efficient inference.Source: E. Implementation Details · p. 5

IV. EXPERIMENTS

The experiments assess GUARD on real HDD scans, ShapeNetPart, and ScanNet, using controlled corruption tests and varied segmentation backbones to examine robustness and generality.

IV. EXPERIMENTS

  • The evaluation asks whether geometric uncertainty identifies corrupted measurements, improves HDD part segmentation, and balances artifact removal, structure preservation, and efficiency.HDD is the application-specific evaluation; ShapeNetPart supports controlled corruption analysis, while ScanNet tests another real-world reconstruction setting.Source: §IV · pp. 5–6

A. Experiment Settings

  • The HDD dataset contains 2,745 scans from different brands and models, split randomly into 80% training and 20% testing.ShapeNetPart provides 12,128 training and 2,874 test samples across 16 categories; its fine-grained object-part focus complements HDD's domain-specific scans.Source: A. Experiment Settings · p. 6
  • ShapeNetPart corruption tests append noisy points at 5%, 10%, 20%, 30%, and 40% severity, matching the observed 5–40% noisy-point range in HDD scans.The protocol appends rather than replaces points because real HDD noise more commonly appears as extra measurements.Source: A. Experiment Settings · p. 6
  • Ghost-cluster corruption clones part points and applies small random rotations and translations, creating globally inconsistent but locally coherent structures.Severity increases the number and displacement of clusters; the experiments use PointNet++ and DGCNN backbones.Source: A. Experiment Settings · p. 6
  • Table I is titled as a baseline comparison under different noise types.Fig. 3 visualizes representative noise.Source: Figure 3 · p. 6, Table I · p. 7
  • ShapeNetPart and HDD inputs are sampled to 2,048 and 2,500 points, respectively, and evaluation reports instance mIoU, retained noise-point ratio, and latency.The random feature dimension is D = 1024; the precision matrix is accumulated online during training and frozen for testing.Source: A. Experiment Settings · p. 7

B. Main Results

GUARD improves segmentation robustness and filtering efficiency across HDD scans and synthetic corruption tests, while experiments also examine a Point Transformer backbone on ScanNetV2.

1) Noise sensitivity of backbones:

  • Under ghost-cluster noise, PointNet++ mIoU falls from 0.8449 clean to 0.6082, while DGCNN falls from 0.8501 to 0.6200.The baseline backbones are evaluated under clean and noisy inputs.Source: 1) Noise sensitivity of backbones · p. 7

2) Comparison with plug-in modules:

  • GUARD improves HDD PointNet++ mIoU by 7.5% over the vanilla backbone and achieves up to 6–8% higher average mIoU on ShapeNetPart under GhostCluster and Local noise.The ShapeNetPart gains are reported even at 40% corruption severity.Source: 2) Comparison with plug-in modules · p. 7

3) Reliability and Efficiency:

  • At S5 GhostCluster severity, GUARD reduces residual noise from 26.9% to 22.3% and also achieves steady gains under Local corruptions.It retains a smaller percentage of noisy points across most corruption types and severity levels.Source: 3) Reliability and Efficiency · p. 7
  • On an RTX 4080, GUARD latency averages 169 ms per PointNet++ sample and 20 ms per DGCNN sample, versus PointCVaR's 623 ms and 26 ms.PointCVaR requires two forward passes; its 95th-percentile latency is 43.5 ms versus GUARD's 27.7 ms.Source: 3) Reliability and Efficiency · p. 7
  • GUARD surpasses PointCVaR across all tested noise types and severity levels, with especially strong reported resilience to ghost-cluster corruption.The comparison also covers local and global random noise.Source: 3) Reliability and Efficiency · p. 7

4) Generalization to Point Transformer:

  • The Point Transformer PTv3 evaluation uses ScanNetV2, a larger and more diverse dataset than ShapeNetPart and HDD, to examine robustness in indoor scenes.The passage identifies Table IV as the location of quantitative results but supplies no values.Source: 4) Generalization to Point Transformer · p. 7

5) Uncertainty Quality:

Geometry-conditioned uncertainty separates corrupted from clean points more clearly than predictive entropy, while GUARD's filtering results and visualizations cover real and synthetic point clouds.

5) Uncertainty Quality:

  • Geometry-conditioned uncertainty σgeo separates noisy from clean points more clearly than predictive entropy, with significantly higher precision under class imbalance.The comparison is shown through ROC and PR curves.Source: 5) Uncertainty Quality · p. 7, Figure 5 · p. 8
  • Manually labeled ScanNet samples are evaluated at a filtering ratio of 0.9 using precision, recall, and F1 for noise detection.ScanNet lacks explicit noise annotations, so a subset of scenes is labeled manually.Source: 5) Uncertainty Quality · p. 7

6) Model Parameter Analysis:

  • GUARD adds approximately 0.4M parameters to both DGCNN and PTv3 while consistently improving denoising quality and downstream segmentation.The reported gains are attributed mainly to uncertainty-guided modeling rather than a substantial increase in model capacity.Source: 6) Model Parameter Analysis · p. 7
  • Fig. 4 evaluates inference latency on an RTX 4080 and segmentation robustness across noise types and severity levels.Table II reports HDD and ScanNet real-scan mIoU separately and ShapeNetPart mIoU at S1 and S5 for GhostCluster, Global, and Local noise.Source: Figure 4 · p. 8, Table II · p. 8

C. Visualization Results

D. Ablation Experiment.

The ablations distinguish geometric uncertainty from semantic entropy and examine contributions within the uncertainty head. Additional ScanNet and model-size comparisons are reported.

D. Ablation Experiment.

  • The GP head provides stronger noise filtering than entropy alone, while entropy alone yields slightly higher mIoU but can over-filter informative regions.GUARD uses geometric uncertainty to detect structural inconsistencies and entropy as a complementary refinement near ambiguous boundaries.Source: D. Ablation Experiment · p. 9
  • On real HDD scans, removing any uncertainty-head component reduces outlier-detection F1, and the full model achieves the best accuracy.MSN improves local robustness, TF enhances global reasoning, MdRFF captures multi-frequency structure, and SN stabilizes GP training.Source: D. Ablation Experiment · p. 9, Table VIII · p. 10
  • ScanNet results include qualitative denoising examples and manually annotated noise-detection performance.Fig. 8 shows removal of floating ghost artifacts and global noise; Table V reports noise-detection performance on manually annotated samples.Source: Figure 8 · p. 10, Table V · p. 10
  • The ablation section also reports a parameter-size comparison in Table VI.The supplied passage identifies the comparison but does not state its values or outcome.Source: Table VI · p. 10

E. Extended Analysis of Robustness and Generalization

The extended analyses examine performance across object categories and the effects of fusion, filtering strength, and uncertainty weighting. They show that the balance between noise removal and feature preservation varies with the design and backbone.

1) Per-Category mIoU Results:

  • ShapeNetPart results improve over vanilla PointNet++ across nearly all categories, with larger gains on geometrically complex or fine-grained objects.The results also report benefits for planar or symmetric categories under high corruption severity.Source: 1) Per-Category mIoU Results · p. 10, Table IX · p. 11
  • Fusion combines the two uncertainty filters in one weighted step and improves segmentation performance while reducing retained noise ratio and mean latency versus Cascade.Table X identifies mIoU, residual noise ratio, and mean latency as the comparison measures.Source: 1) Per-Category mIoU Results · p. 10, Table X · p. 11

2) Additional Analysis on DGCNN on HDD Data:

  • DGCNN + GUARD on HDD data scores below vanilla DGCNN but still outperforms PointCVaR; aggressive filtering can remove informative points used by EdgeConv.The filtering-ratio analysis reports a better balance between noise suppression and feature preservation at a moderate ratio.Source: 2) Additional Analysis on DGCNN on HDD… · pp. 10–11, Table XI · p. 11
  • Lower α emphasizes entropy and slightly raises mIoU while retaining more noisy points; higher α emphasizes geometric uncertainty and removes more fake points, sometimes with a small mIoU reduction.Table XII fixes the filtering ratio at 0.9 for this fusion-weight comparison.Source: 2) Additional Analysis on DGCNN on HDD… · p. 11, Table XII · p. 11

V. DISCUSSION AND ENGINEERING IMPLICATIONS

The discussion distinguishes geometric reliability from semantic confidence and emphasizes the tradeoff between suppressing artifacts and preserving useful geometry. The conclusion frames GUARD as a perception-level contribution, while downstream manipulation and disassembly performance remain future work.

V. DISCUSSION AND ENGINEERING IMPLICATIONS

  • A corrupted structure can remain locally plausible and receive a confident semantic prediction even when its geometry is inconsistent with the object.The uncertainty analyses indicate that geometric uncertainty provides information complementary to predictive entropy, particularly for structured reconstruction artifacts.Source: §V · p. 11
  • More aggressive filtering can remove additional corrupted measurements but may also discard informative points, so segmentation accuracy alone does not characterize denoising quality.Filtering strategy should consider both retained-geometry reliability and usefulness for component identification.Source: §V · p. 11

VI. CONCLUSION

  • The experiments evaluate perception-level filtering, segmentation, and corruption detection across HDD, ShapeNetPart, and ScanNet; downstream robotic manipulation and disassembly performance remains future work.The conclusion presents GUARD as a perception-level foundation for component handling under imperfect sensing conditions.Source: §V · pp. 11–12, §VI · p. 12

Related papers

Loading 2610.09068v1…