Deep Learning For Point Cloud Denoising: A Survey

Chengwei Zhang, Xueyi Zhang, Mingrui Lao, Tao Jiang, Xinhao Xu, Wenjie Li, Fubo Zhang, Longyong Chen · 2025

Read Deep Learning For Point Cloud Denoising: A Survey by Chengwei Zhang, Xueyi Zhang, Mingrui Lao, Tao Jiang, Xinhao Xu, Wenjie Li, Fubo Zhang, Longyong Chen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-11932.txt.

Introduction

Chengwei Zhang and co-authors survey deep learning methods that clean point clouds in a 2025 paper. Point clouds from real scenes always carry noise, in varied forms and strengths. Cleaning them first lifts every later task, from mapping to object spotting.

Learning-based cleaners now beat classic methods, thanks to flexible designs and strong pattern skills. Yet no full survey had organized this fast-growing shelf of work. The authors frame cleaning as two steps: throwing out stray points, then smoothing the rest. They sort methods into a taxonomy built for the task and name the open challenges. Newcomers get a map, and veterans get a list of unsolved problems.

Catalog record

CREATORChengwei Zhang, Xueyi Zhang, Mingrui Lao, Tao Jiang, Xinhao Xu, Wenjie Li, Fubo Zhang, Longyong Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-11932

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Deep Learning For Point Cloud Denoising: A Survey Chengwei Zhang, Xueyi Zhang, Mingrui Lao, Tao Jiang, Xinhao Xu, Wenjie Li, Fubo Zhang, Longyong Chen Abstract Real-world environment-derived point clouds invariably exhibit noise across varying modalities and intensities. Hence, point cloud denoising (PCD) is essential as a preprocessing step to improve downstream task performance. Deep learning (DL)-based PCD models, known for their strong representation capabilities and flexible architectures, have surpassed traditional methods in denoising performance. To our best knowledge, despite recent advances in performance, no comprehensive survey systematically summarizes the developments of DL-based PCD. To fill the gap, this paper seeks to identify key challenges in DL-based PCD, summarizes the main contributions of existing methods, and proposes a taxonomy tailored to denoising tasks. To achieve this goal, we formulate PCD as a two-step process: outlier removal and surface noise restoration, encompassing most scenarios and requirements of PCD. Additionally, we compare methods in terms of similarities, differences, and respective advantages. Finally, we discuss research limitations and future directions, offering insights for further advancements in PCD. 1 Introduction Point clouds serve as a fundamental 3D data format, bridging the gap between depth sensors and 3D vision algorithms. However, real-world point clouds inevitably contain noise and artifacts, which degrade the performance of downstream tasks such as point cloud classification Sun et al. 2022 , object detection Zhang et al. 2024b , semantic segmentation Yan et al. 2024 , and surface reconstruction Huang et al. 2024 . Compared to traditional geometric methods, deep learning (DL)-based point cloud denoising (PCD) has gained significant attention in the research community due to its superior denoising quality and generalization capability to improve the performance of downstream tasks Zhao et al. 2024 ; Li et al. 2024 ; Zhang et al. 2024a ; de Silva Edirimuni et al. 2024 ; Mao et al. 2024 . The primary causes and foundmental type of noisy points is illustrated in the upper part of Fig. 1 . On the one hand, the sensing imperfections and misalignment in multi-view scanning resulting points being distributed in a non-uniform manner with varying thickness and blurry geometric details, leading to surface noise Huang et al. 2024 . Moreover, in adverse weather conditions, active pulsed systems such as LiDAR are particularly vulnerable in scattering media such as rain, snow, and fog. Particles in these media interact with the laser beam resulting spurious returns at incorrect distances, introducing outliers Zhao et al. 2024 . On the other hand, noise in point clouds can also be intentionally introduced with malicious intent, such as through delicately designed adversarial outliers or slight perturbations of origional surface points Li et al. 2024 ; Sun et al. 2023 . To effectively mitigate noise, existing DL-based PCD methods can be broadly categorized into consecutive and interrelated steps: outlier removal and surface noise restoration , as illustrated in the lower part of Fig. 1 , which together address most scenarios and requirements of PCD. Specifically, outlier removal aims to drop points to enhance the stability of feature extraction , and, as subsequent, surface noise restoration is designed to move points to accurately and uniformaly relocate it onto the underlying surface while faithfully preserve fine-grained edge geometric details . Figure 1: Illustration of the primary causes of noisy points and the denoising process. The color bar represents the point distribution density, where darker indicate higher density. The point colors range from red to blue , representing the distance to the underlying surface. However, existing surveys Chen et al. 2024b ; Sohail et al. 2024 only focus on certain aspects of DL-based PCD. Chen et al. 2024b adopts a task-oriented perspective, primarily focusing on the graph neural network-based architectures design for surface noise restoration. Sohail et al. 2024 treats surface noise restoration as a subtask of point cloud enhancement, covers a limited part of surface noise restoration methods and omits the review of outlier removal methods. Moreover, their taxonomy lacks consideration of the intrinsic characteristics of denoising tasks, with insufficent discussion on the interrelations between different methods. To fill this gap, this survey aims to provide a comprehensive review of DL-based PCD methods. Specifically, we consider the holistic process of outlier removal and surface noise restoration and present a generalized pipeline for DL-based PCD, covering data formats, denoising processes, and learning paradigms. Additionally, we provide a detailed overview of datasets and evaluation metrics for DL-based PCD. Moreover, we propose a novel taxonomy for DL-based PCD, organizing existing methods in a top-down manner based on key challenges. Spercifically, this taxonomy includes two primary categories: outlier removal and surface noise restoration, with the latter further divided into edge-aware denoising and precise surface recovery. Additionally, we summarize key technical contributions from existing methods, forming a fine-grained, bottom-up classification of specific solutions. Further more, we provide a comparative analysis of their similarities, differences, and respective advantages and disadvantages. Finally, we evaluate representative methods under a unified benchmark setting to assess their denoising performance. Based on our findings, we outline potential future research directions in DL-based PCD, including multi-task learning and improving noise generalization. We summarize the main contributions as follows: • To the best of our knowledge, this is the first work that systematically and comprehensively reviews DL-based PCD methods, covering two fundamental stages: outlier removal and surface noise restoration, which encompassing most scenarios and requirements of PCD. • Guided by the requirements and characteristics of PCD, we propose a taxonomy to categorize existing methods, as illustrated in Fig. 2 . Furthermore, we analyze the intrinsic connections between the two denoising stages and explore the underlying relationships among different approaches. • We evaluate representative DL-based PCD methods in the most frequently used settings. In addition, we discuss open challenges and promising future directions to inspire further research in this field. Deep Learning on Point Cloud Denoising Outlier Removal (Section 3 ) Point-based (Section 3.1 ) PointCVaR Li et al. 2024 , TripleMixer Zhao et al. 2024 Range-Image-based (Section 3.2 ) WeatherNet Heinzler et al. 2020 , Slide Bae et al. 2022 Surface Noise Removal (Section 4 ) Edge-Aware Denoising (Section 4.1 ) Geometry Prior GeoDualCNN Wei et al. 2021 , DFP Lu et al. 2020 Position-Normal Interaction PCDNF Liu et al. 2023 , PN-Internet Yi et al. 2024 Supervision Strategy PointFilter Zhang et al. 2020 , Re-PCD Chen et al. 2022b Precise Surface Recovery (Section 4.2 ) Denoising Path Learning Path-Net Wei et al. 2024 , StraightPCF de Silva Edirimuni et al. 2024 Surface Consistency N2NM Ma et al. 2023 , SVCNet Zhao et al. 2022 Surface Field Gradient Field: Score Luo and Hu 2021 , PSR Chen et al. 2022a Diffusion Field: P2P-Bridge Vogel et al. 2025 Neural Implicit Field: LIFPCF Wang et al. 2024a , N2NM Ma et al. 2023 Manifold Reconstruction Manifold Sampling: DMR Luo and Hu 2020 , SSPCN Li and Sheng 2023 Manifold Refinement: PD-LTS Mao et al. 2024 , PD-Refiner Zhang et al. 2024a Figure 2: A taxonomy of representative DL-based PCD methods. Figure 3: An overview of DL-based PCD paradigms. The outlier denoiser can be regarded as a point-wise classification network that removes outliers based on probability estimation. The surface noise denoiser takes the outlier-denoised point cloud as input and follows either point-level or patch-level

Read the complete text