Texas A&M University / 3D perception

GUARD.Geometric Uncertainty-Aware
Robust Denoiser for
Point Cloud Segmentation

Reliable geometry. Robust segmentation.
Unreliable-point filtering and part understanding in a single forward pass.

Zuoxu WangXiao Liang*

Zachry Department of Civil & Environmental Engineering
Texas A&M University

* Corresponding author

Read abstractCode PublicBibTeX
From noisy scans to reliable parts
OriginalOriginal HDD scan 1 with noisy and duplicated pointsNoisy point cloud
PointCVaRPointCVaR filtering result for HDD scan 1Comparison method
GUARD OursGUARD filtering and segmentation result for HDD scan 1Geometry-aware filtering
Real HDD point clouds from the paper. GUARD identifies geometric inconsistencies to remove scanning artifacts while retaining meaningful part structure.
2,745Real HDD point clouds
77.39 → 83.18%HDD mIoU · PointNet++
0.7931ScanNet noise-detection F1
Single passJoint filtering & segmentation
01 / The problem

Reliable perception starts
with reliable geometry.

Ghost structures can look locally plausible while being globally inconsistent. GUARD learns which geometric representations to trust.

Abstract

Reliable part segmentation of scanned hard disk drives (HDDs) is an important perception requirement for automated component handling and robotic disassembly in e-waste recycling. Although point clouds provide detailed geometric information for such tasks, real scans often contain sensor noise and reconstruction artifacts, particularly ghost structures that are locally plausible yet geometrically inconsistent, making both noise removal and semantic interpretation unreliable.

We propose 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.

GUARD is primarily evaluated on 2,745 real HDD point clouds containing scanning and reconstruction artifacts. On this dataset, it improves PointNet++ segmentation mIoU from 0.7739 to 0.8318 and demonstrates robust performance across different backbones with limited parameter overhead and efficient single-pass inference. Further evaluation on ShapeNetPart and ScanNet assesses robustness and generality beyond the HDD setting, with geometric uncertainty achieving an F1 score of 0.7931 for corrupted-point detection on manually annotated ScanNet samples.

These results support the potential of geometric reliability modeling to provide robust perception for automated robotic disassembly in manufacturing and other engineering applications involving noisy point-cloud data.

02 / Method

Geometry and uncertainty,
working together.

A lightweight module for PointNet++, DGCNN, and Point Transformer backbones.

GUARD architecture: multi-scale geometric encoding and a spectrally normalized transformer feed a Gaussian Process uncertainty head; fusion with predictive entropy guides point filtering and segmentation.
GUARD framework. A multi-scale geometric encoder, geometry-aware transformer, and Gaussian Process head estimate point-wise reliability. Geometric uncertainty and predictive entropy guide the final filtering. Select the figure to view it at full resolution.
01

Encode geometry

Multi-scale neighborhoods capture fine and coarse structure. A geometry-aware transformer connects local features with global context.

02

Estimate reliability

A multi-bandwidth random Fourier feature Gaussian Process measures uncertainty in geometric embeddings. Spectral normalization stabilizes learning.

03

Filter and segment

Geometric uncertainty detects structural inconsistencies; predictive entropy refines ambiguous boundaries within the same forward pass.

03 / Quantitative results

Robust across
corrupted geometry.

Real scans and controlled corruptions, evaluated with the same segmentation backbones.

Segmentation benchmarks

mIoU ↑ · higher is better

Bar axis: 0–1 mIoU. ShapeNetPart average is the paper’s reported average over synthetic corruptions, not an average across HDD and ScanNet.

Full benchmark table All reported noise types and severity endpoints
Robustness under different noise types (manuscript Table 2). S1 = 5%; S5 = 40%. Values are mIoU.
BackboneMethodHDDScanNetGhost S1Ghost S5Global S1Global S5Local S1Local S5Avg. GhostAvg. GlobalAvg. LocalAvg. All

Generalization to Point Transformer

ScanNet · Point Transformer V3

MethodmIoU ↑
PTv30.7757
PTv3 + GUARD0.7902

A 1.45 percentage-point gain on ScanNet, with approximately 0.4M additional parameters.

Modest parameter overhead

Base parameters + GUARD addition, in millions

BackboneBaseAdded
PointNet++2.2M+0.3M
DGCNN1.9M+0.4M
Point Transformer46.2M+0.4M
Paper figure comparing GUARD and PointCVaR inference latency on PointNet++ and DGCNN.
Efficient inference. The paper reports approximately 169 ms for PointNet++ + GUARD and 20 ms for DGCNN + GUARD on an RTX 4080.
Paper figure showing segmentation robustness across noise severities and corruption types.
Robustness under corruption. Segmentation performance across the paper’s noise types and severity levels.
04 / Qualitative results

See what GUARD keeps.

Original scans, PointCVaR, and GUARD outputs from the manuscript.

Rows compare methods. Columns show different objects.

From object parts to indoor scenes

ScanNet reconstructions contain floating and duplicated artifacts at room scale. Highlighted regions show corruption in the original input; the corresponding filtered scene appears on the right.

ScanNet scene 0019: original reconstruction with highlighted artifacts on the left and GUARD result on the right.
ScanNet · Scene 0019
ScanNet scene 0046, comparing the original scene and GUARD filtered reconstruction.
ScanNet · Scene 0046
05 / Uncertainty quality

Detect geometric unreliability.

Geometric uncertainty separates corrupted points more effectively than predictive entropy on the annotated ScanNet samples.

Corrupted-point detection

Manually annotated ScanNet samples · filtering ratio 0.9

MeasurePrecision ↑Recall ↑F1 ↑
Predictive entropy0.22840.21440.2212
GUARD σgeo0.80320.78320.7931
Geometric uncertainty0.7931

F1 score for corrupted-point detection
on the annotated ScanNet evaluation.

ROC curves from the manuscript comparing geometric uncertainty and predictive entropy for noise detection.
Receiver operating characteristic
Precision-recall curves from the manuscript comparing geometric uncertainty and predictive entropy.
Precision–recall
06 / Ablations & trade-offs

What makes the
uncertainty head work?

Component ablations on real HDD scans and a comparison of geometric uncertainty with entropy.

Uncertainty-head components

HDD outlier detection · higher is better

VariantPrecisionRecallF1
Baseline (RFF)0.33580.25370.2890
Without MSN0.97710.09530.1736
Without TF0.71920.59790.6530
Without MdRFF0.28510.86360.4287
Without SN0.21080.13240.1626
GUARD (full)0.83300.58970.6906

MSN: multi-scale neighborhoods; TF: transformer; MdRFF: multi-bandwidth random Fourier features; SN: spectral normalization.

Geometry vs. semantic ambiguity

ShapeNetPart · uncertainty-component ablation

ModuleNoise kept ↓mIoU ↑
Only entropy16.1%0.7203
Only GGPH11.7%0.6672

Geometric uncertainty provides stronger filtering, while entropy alone yields higher segmentation mIoU in this ablation.

A filtering trade-off on HDD

DGCNN + GUARD scores 0.7425 mIoU versus 0.7434 for vanilla DGCNN. The paper examines how aggressive filtering can remove informative points used by local graph construction.

Residual noise by corruption severity Lower noise kept is better
Average noise kept (%), exactly as reported in the manuscript.
CorruptionMethodS1S2S3S4S5
LocalPointCVaR4.99.017.223.930.0
LocalGUARD4.48.114.620.223.7
GlobalPointCVaR2.44.68.411.012.6
GlobalGUARD1.52.96.411.919.8
GhostClusterPointCVaR5.78.09.922.826.9
GhostClusterGUARD5.57.38.920.122.3

GUARD retains more global-noise points than PointCVaR at S4 and S5; the complete table preserves this trade-off.

07 / Resources

Build on GUARD.

The repository contains implementation, evaluation, training, and visualization instructions. The code repository is public; the full HDD dataset release is planned after publication.

GUARD_Point_denoiser on GitHub

BibTeX

Manuscript citation; publication venue and DOI are not specified in the supplied source.

@unpublished{wang_guard,
  title  = {{GUARD}: Geometric Uncertainty-Aware Robust
            Denoiser for Point Cloud Segmentation},
  author = {Wang, Zuoxu and Liang, Xiao},
  note   = {Manuscript},
  url    = {https://github.com/001-Wang/GUARD_Point_denoiser}
}