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

Zuoxu Wang; Xiao Liang

cs.CVOct 6, 2026

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

Authors: Zuoxu Wang, Xiao Liang

Organizations: Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843 USA

Abstract

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/.

Figures & tables

Fig. 1: Overall architecture of the proposed framework. Our method introduces an outlier-removal head composed of a multi-scale geometric encoder, a geometry-aware transformer block, and an uncertainty-aware Gaussian Process (GP) head. As illustrated, it can be seamlessly attached to various backbones (e.g., PointNet++, DGCNN), providing a unified plug-in module for robust and uncertainty-aware point cloud segmentation.

Panel of Figure 2 — Fig. 2: Motivation for geometric uncertainty modeling. (a) We hypothesize that clean points lie near a compact geometric manifold in the embedding space, while outlier or ghost-cluster points deviate from this manifold, leading to higher predictive uncertainty. (b) The predicted variance $\sigma^{2}$ distribution on the HDD dataset supports this hypothesis: most samples concentrate in a low-uncertainty region, while deviations correspond to abnormal or noisy structures.

Panel of Figure 3 — Fig. 3: Visualization of representative noise.

Backbone	Clean	Ghost	Drop (%)	Global	Drop (%)	Local	Drop (%)
PointNet++	0.8449	0.6082	$-28.0$	0.3242	$-61.6$	0.5073	$-40.0$
DGCNN	0.8501	0.6200	$-27.1$	0.4310	$-49.3$	0.5582	$-34.3$

TABLE I: Baseline comparison under different noise types.

Backbone	Method	HDD mIoU $\uparrow$	ScanNet mIoU $\uparrow$	ShapeNetPart (synthetic corruptions; S1=5%, S5=40%)
Backbone	Method	HDD mIoU $\uparrow$	ScanNet mIoU $\uparrow$	GhostCluster mIoU $\uparrow$		Global mIoU $\uparrow$		Local mIoU $\uparrow$		Averages over ShapeNet only $\uparrow$
				S1	S5	S1	S5	S1	S5	mIoU GC	mIoU G	mIoU L	mIoU A
PointNet++	Vanilla	0.7739	0.5259	0.8043	0.6082	0.6543	0.3242	0.8091	0.5073	0.7175	0.4729	0.6757	0.6221
	+PointCVaR	0.6685	0.5281	0.7840	0.6049	0.7244	0.4125	0.7851	0.5207	0.7049	0.5699	0.6672	0.6473
	+ GUARD	0.8318	0.5601	0.8176	0.7014	0.8111	0.4409	0.8206	0.6131	0.7670	0.6396	0.7220	0.7095
DGCNN	Vanilla	0.7434	0.4865	0.7878	0.6200	0.7421	0.4310	0.7902	0.5582	0.7139	0.5879	0.6805	0.6608

TABLE II: Robustness under different noise types. HDD/ScanNet (real) mIoU is reported separately. The ShapeNet block reports mIoU at 5% (S1) and 40% (S5) severity for GhostCluster , Global , and Local noise, plus averages (mIoU GC , mIoU G , mIoU L , overall mIoU A ).

Type	Method	s1 $\downarrow$	s2 $\downarrow$	s3 $\downarrow$	s4 $\downarrow$	s5 $\downarrow$
Local	PointCVaR	4.9	9.0	17.2	23.9	30.0
	GUARD	4.4	8.1	14.6	20.2	23.7
Global	PointCVaR	2.4	4.6	8.4	11.0	12.6
	GUARD	1.5	2.9	6.4	11.9	19.8
Ghostcluster	PointCVaR	5.7	8.0	9.9	22.8	26.9
	GUARD	5.5	7.3	8.9	20.1	22.3

TABLE III: Average noise kept (%) under different corruption types.

Panel of Figure 7 — Fig. 4: Efficiency and robustness evaluation of the proposed method. (a) Inference latency comparison across different methods on an RTX 4080 GPU. (b) Segmentation robustness under varying noise types and severity levels.

Method	mIoU	Wall	Floor	Cab.	Bed	Chair	Sofa	Table	Door	Window	Shelf
PTv3	0.7757	0.8724	0.9568	0.7129	0.8343	0.9294	0.8623	0.7912	0.7499	0.7390	0.8302
+ GUARD	0.7902	0.8785	0.9609	0.7208	0.8384	0.9372	0.8721	0.8043	0.7520	0.7445	0.8364
Method	mIoU	Picture	Counter	Desk	Curtain	Fridge	Shower	Toilet	Sink	Bathtub	OtherFurn.
PTv3	0.7757	0.3979	0.7168	0.7403	0.7896	0.7119	0.7244	0.9336	0.7095	0.8821	0.6292
+ GUARD	0.7902	0.4018	0.7325	0.7421	0.7996	0.7043	0.7227	0.9614	0.6923	0.8896	0.6338

TABLE IV: Per-category mIoU results on ScanNet for PointTransformerV3 (PTv3) and our GUARD. Improved results are in bold.

Panel of Figure 9 — Fig. 5: PR and ROC curves for noise detection.

Method	Precision $\uparrow$	Recall $\uparrow$	F1 Score $\uparrow$
Predictive Entropy (PE)	0.2284	0.2144	0.2212
GUARD ( $\sigma_{\text{geo}}$ )	0.8032	0.7832	0.7931

TABLE V: Noise detection performance on manually annotated ScanNet samples.

Model	DGCNN	PointNet++	Point Transformer
# Params (M)	1.9(+0.4)	2.2(+0.3)	46.2(+0.4)

TABLE VI: Parameter size (M) comparison.

Panel of Figure 12 — Fig. 6: Visualization of segmentation results on real HDD point clouds.

Panel of Figure 13 — Fig. 7: Visualization of segmentation results on representative ShapeNetPart categories under noisy conditions.

Panel of Figure 14 — Fig. 8: Qualitative results on ScanNet. Our method effectively removes floating ghost artifacts and global noise, yielding cleaner reconstructions in complex indoor scenes.

Module	Avg. Noise Kept(%) $\downarrow$	Instance mIoU $\uparrow$
Only Entropy	16.1	0.7203
Only GGPH	11.7	0.6672

TABLE VII: GGPH vs. entropy ablation on ShapeNetPart.

Method	Precision	Recall	F1
Baseline (RFF)	0.3358	0.2537	0.2890
GUARD w/o MSN	0.9771	0.0953	0.1736
GUARD w/o TF	0.7192	0.5979	0.6530
GUARD w/o MdRFF	0.2851	0.8636	0.4287
GUARD w/o SN	0.2108	0.1324	0.1626
GUARD (full)	0.8330	0.5897	0.6906

TABLE VIII: Each component contributes to robustness.

Backbone	Method	Avg. mIoU $\uparrow$	Per-Category mIoU $\uparrow$ on ShapeNetPart
Backbone	Method	Avg. mIoU $\uparrow$	Earphone	Motorbike	Rocket	Car	Laptop	Cap	Skateboard	Mug	Guitar	Bag	Lamp	Table	Airplane	Pistol	Chair	Knife
PointNet++	Vanilla	0.6082	0.4786	0.5345	0.4609	0.5102	0.7431	0.5680	0.4366	0.7016	0.7709	0.5754	0.7400	0.5774	0.6345	0.6380	0.6991	0.6624
	+PointCVaR	0.6049	0.5242	0.3726	0.4437	0.5739	0.7984	0.5998	0.4594	0.6300	0.7889	0.4863	0.7091	0.6374	0.6656	0.6378	0.7057	0.6456
	+ GUARD	0.7014	0.6363	0.5841	0.4977	0.5836	0.8529	0.7008	0.6025	0.7126	0.8786	0.5645	0.7939	0.7246	0.7741	0.6660	0.8394	0.8108
DGCNN	Vanilla	0.6200	0.5847	0.3857	0.4199	0.5224	0.7648	0.6289	0.4293	0.6433	0.7915	0.6242	0.7600	0.6664	0.6700	0.5868	0.7725	0.6696
	+PointCVaR	0.6555	0.6501	0.4082	0.4530	0.5209	0.8413	0.6993	0.4469	0.5840	0.8265	0.6831	0.7667	0.7124	0.7137	0.6307	0.8103	0.7409

TABLE IX: Per-category mIoU results on ShapeNetPart(ghostcluster-s5) for different methods. We report results for PointNet++ and DGCNN backbones under three settings: Vanilla, PointCVaR, and our GUARD.

Method	mIoU $\uparrow$	Noise Ratio (%) $\downarrow$	Mean Latency (ms) $\downarrow$
Cascade	0.6282	24.41	189.68
Fusion (ours)	0.7320	22.30	169.51

TABLE X: Comparison between Fusion and Cascade strategies. mIoU measures segmentation accuracy, Noise Ratio (%) denotes the proportion of residual noisy points, and mean latency (ms) reflects average inference time per sample.

Filtering Ratio	0.70	0.75	0.80	0.85	0.90	0.95
mIoU $\uparrow$	0.7212	0.7295	0.7341	0.7373	0.7425	0.7406

TABLE XI: Effect of filtering ratio for GUARD when applied to DGCNN on the HDD dataset. A moderate filtering ratio achieves the best balance between noise suppression and feature preservation.

$\alpha$	0.1	0.3	0.5	0.7	0.9
Noise Kept (%) $\downarrow$	18.8	13.5	9.9	7.8	6.7
mIoU $\uparrow$	0.7834	0.7702	0.7571	0.7425	0.7301

TABLE XII: Effect of the fusion weight

\alpha

between predictive entropy and geometric uncertainty

\sigma^{2}

on the HDD dataset using DGCNN. Lower

\alpha

improves mIoU but retains more noisy points. The filtering ratio is fixed to 0.9.

Explore similar work

UGD: An Unsupervised Geometric Distance for Evaluating Real-world Noisy Point Cloud Denoising
Apr 18, 2026Zhiyong Su, Jincan Wu, Yonghui Liu +2Point Clouds Distance
School of Automation, Nanjing University of Science and Technology, Nanjing 210094, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
Learning Point Cloud Geometry as a Statistical Manifold: Theory and Practice
May 11, 2026Jinwoo Lee, Jiwoo Kim, Woojae Shin +2Point Clouds Robust Geometric Model Estimation
Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea · Daegu Gyeongbuk Institute of Science and Technology (DGIST), Daegu, Republic of Korea
RODR: Riemannian Orthogonally Decoupled Regularization for Disentangled Manifold Representation
Jul 27, 2026Jiayu Zhu, Wenlai ZhaoPoint Clouds Riemannian Manifolds
Department of Computer Science and Technology, Tsinghua University, Beijing, 100084, China

Apr 18, 2026, cs.CVJ/K move · Enter open · S save

UGD: An Unsupervised Geometric Distance for Evaluating Real-world Noisy Point Cloud Denoising

Zhiyong Su, Jincan Wu, Yonghui Liu, Zheng Li+1

School of Automation, Nanjing University of Science and Technology, Nanjing 210094, China · School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China

Point Clouds Distance

Open paper AskPDF arXiv

Point cloud denoising is a fundamental and crucial challenge in real-world point cloud applications. Existing quantitative evaluation metrics for point cloud denoising methods are implemented in a supervised manner, which requires both the denoised point cloud and the corresponding ground-truth clean point cloud to compute a representative geometric distance. This requirement is highly problematic in real-world scenarios, where ground-truth clean point clouds are often unavailable. In this paper, we propose a simple yet effective unsupervised geometric distance (UGD) for real-world noisy point cloud denoising, calculated solely from noisy point clouds. The core idea of UGD is to learn a patch-wise prior model from a set of clean point clouds and then employ this prior model as the ground-truth to quantify the degradation by measuring the geometric variations of the denoised point cloud. To this end, we first learn a pristine Gaussian Mixture Model (GMM) with extracted patch-wise quality-aware features from a set of pristine clean point clouds by a patch-wise feature extraction network, which serves as the ground-truth for the quantitative evaluation. Then, the UGD is defined as the weighted sum of distances between each patch of the denoised point cloud and the learned pristine GMM model in the patch space. To train the employed patch-wise feature extraction network, we propose a self-supervised training framework through multi-task learning, which includes pair-wise quality ranking, distortion classification, and distortion distribution prediction. Quantitative experiments with synthetic noise confirm that the proposed UGD achieves comparable performance to supervised full-reference metrics. Moreover, experimental results on real-world data demonstrate that the proposed UGD enables unsupervised evaluation of point cloud denoising methods based exclusively on noisy point clouds.

Abstract

Explore similar work

UGD: An Unsupervised Geometric Distance for Evaluating Real-world Noisy Point Cloud Denoising

Learning Point Cloud Geometry as a Statistical Manifold: Theory and Practice

RODR: Riemannian Orthogonally Decoupled Regularization for Disentangled Manifold Representation