PaperlayerView exact revision on arXiv

Source-linked AI summary

Advancements in Point Cloud Data Augmentation for Deep Learning: A Survey

Qinfeng Zhu, Lei Fan, Ningxin Weng

arXiv:2308.12113v5cs.CV

TL;DR

Point-cloud augmentation is widely used, but systematic surveys are lacking, and point clouds have distinct characteristics and limited public dataset diversity. This survey organizes methods into basic and specialized categories and reports that augmentation improves performance in typical downstream tasks.

  • Problem

    Point-cloud augmentation methods are diverse, but systematic surveys are lacking, while point clouds have distinct characteristics and limited public benchmark data diversity.

    Source: §1 · p. 2

  • Method

    The paper surveys point-cloud augmentation methods, organizes them into basic and specialized categories, and discusses their strengths and limitations.

    Source: §1 · pp. 2–3, §5 · p. 32

  • Results

    The survey reports that augmentation improves deep-learning performance in typical downstream tasks, while basic methods remain prevalent and specialized methods serve task-specific needs.

    Source: §4 · p. 29, §5 · pp. 32–33

  • Takeaways & Limitations

    Basic methods are often preferred for their simplicity and adaptability, while specialized methods tend to be more complex and task-specific.

    Source: §5 · pp. 32–33

    The survey cannot quantitatively compare augmentation methods across studies because datasets, networks, and experimental setups are inconsistent.

    Source: §5 · p. 33

Abstract

from arXiv · show

Deep learning (DL) has become one of the mainstream and effective methods for point cloud analysis tasks such as detection, segmentation and classification. To reduce overfitting during training DL models and improve model performance especially when the amount and/or diversity of training data are limited, augmentation is often crucial. Although various point cloud data augmentation methods have been widely used in different point cloud processing tasks, there are currently no published systematic surveys or reviews of these methods. Therefore, this article surveys these methods, categorizing them into a taxonomy framework that comprises basic and specialized point cloud data augmentation methods. Through a comprehensive evaluation of these augmentation methods, this article identifies their potentials and limitations, serving as a useful reference for choosing appropriate augmentation methods. In addition, potential directions for future research are recommended. This survey contributes to providing a holistic overview of the current state of point cloud data augmentation, promoting its wider application and development.

1 Introduction

Point-cloud augmentation methods are diverse, and point clouds differ from images while public datasets often have limited labels and diversity. This article surveys the methods and organizes them into basic and specialized categories.

1 Introduction

  • Data augmentation can increase training-data quantity and diversity and is used to improve robustness, generalization, and reduce overfitting.Its role is especially relevant when training datasets are limited.Source: §1 · p. 2
  • Point clouds’ distinct characteristics and limited public benchmarks motivate tailored augmentation methods rather than assuming image methods transfer unchanged.Image operations such as rotation and scaling can be applied, but their performance may vary between 2D images and 3D point clouds.Source: §1 · p. 2
  • Existing intra-domain and inter-domain labels do not capture the diversity of point-cloud augmentation methods, complicating method selection.The article argues that systematically surveying and categorizing the methods is valuable.Source: §1 · p. 2
  • The proposed taxonomy divides methods into basic augmentation, which is simple and versatile, and specialized augmentation, which targets particular challenges or contexts and is typically more complex.The complexity of specialized methods depends on implementation details.Source: §1 · pp. 2–3
  • The article presents a comprehensive survey of point-cloud augmentation methods, discusses their applications, and suggests potential future research directions.The authors describe it as the first comprehensive survey of these methods to their knowledge.Source: §1 · pp. 3–4

2 Basic point cloud augmentation

Basic point-cloud augmentation diversifies training data through geometric transforms, point removal, jittering, instance insertion, global or local operations, and automated operation selection. Its use must account for task and scene context, semantic plausibility, preservation of important structure, and the risk that training-tailored augmentations may undermine generalization to unseen data.

2.1.1 Affine transformation

  • Affine transformations—translation, rotation, scaling, and flipping—vary instance location, pose, size, and orientation, but must avoid unrealistic geometry or semantic placements.Translation and rotation ranges should avoid occlusions, inappropriate locations, information loss, and distortion; scaling should preserve realistic geometry and point density. Upright-instance datasets may need rotations for tilt, while vertical flipping can create semantic ambiguity in scene-level data.Source: §2.1.1 · pp. 4–6

2.1.2 Drop

  • Drop augmentation improves robustness to missing or incomplete scans, but excessive or poorly targeted removal can erase key information and impair learning.Strategies range from random point removal and cuboid- or circle-based erasing to adversarial key-point dropping and selective removal that preserves plant edges.Source: §2.1.2 · pp. 6–7

2.1.3 Jittering

  • Jittering adds pointwise positional noise to simulate measurement variation and improve noise robustness, but excessive perturbation can damage local structure.It is frequently used for segmentation and classification but rarely for detection, where precise spatial relationships are important; task-specific testing is needed.Source: §2.1.3 · p. 8

2.1.4 GT-sampling

  • GT-sampling adds labeled instances, mainly in scene-level data, but placement can be semantically implausible and fail to reproduce realistic occlusion.Methods seek reasonable placement, address class imbalance by adding minority-class instances, or model occlusion by retaining nearer LiDAR points; combining semantic plausibility with balancing remains beneficial.Source: §2.1.4 · pp. 8–10

2.2 Global and local point cloud augmentation

  • Global augmentation transforms an entire scene, whereas local and patch-based augmentation modify selected instances or parts to improve learning of local features.Patch strategies include partitioning instances into pyramids or cuboids, swapping parts, and applying local deformations or grouped basic operations; some approaches may not suit particular object classes.Source: §2.2 · pp. 10–11

2.3 Application of basic operations

  • Basic operations are commonly combined by task, but their best mix is often unclear and must be found through extensive testing.GT-sampling is frequent in detection, while classification commonly combines translation, rotation, scaling, jittering, and drop; rotation and drop may also be used independently. Auto-optimization methods learn operation combinations or parameters, though classification-oriented methods need further validation for other tasks.Source: §2.3 · pp. 11–13
  • LidarAugment uses factorization and alignment search to reduce hyperparameters and computational complexity while selecting schemes among ten operations.Source: §2.3 · pp. 13–14
  • Auto-optimization may select training-tailored augmentations that diverge from unseen data and weaken generalization.Future work could assess augmented-data generalizability or restrict operations to reduce deviations from real-world scenarios.Source: §2.3 · p. 14

3 Specialized point cloud augmentation

Specialized point cloud augmentation targets specific challenges or application contexts. The survey finds distinct task patterns across methods, including mixup for classification and domain augmentation for outdoor detection.

3 Specialized point cloud augmentation

  • Specialized methods span varied tasks and settings; mixup is most common for classification, while domain augmentation is predominantly used for outdoor detection.Specialized methods are typically tailored to specific challenges or application contexts; segmentation uses them comparatively less, usually multi-modal, adversarial deformation, or domain techniques.Source: §3 · pp. 14–15, Table 2 · p. 14

3.1 Mixup augmentation

  • Mixup combines two point cloud instances, but methods trade structural preservation against boundary continuity and may damage semantic information.PointMixup can damage structure; RSMix preserves shape but produces boundary discontinuities, while SageMix safeguards significant local areas; mixup also tends to limit adaptation beyond classification.Source: §3.1 · pp. 16–17

3.2 Domain augmentation

  • Domain augmentation simulates other environments or sensors in training data to address domain gaps between training and testing point clouds.Examples include modeling fog, snow, rain, or sensor characteristics; snowfall simulation modifies beam measurements and models ground moisture.Source: §3.2 · pp. 18–19

3.3 Adversarial deformation augmentation

The survey describes specialized approaches that deform, densify, complete, generate, or enrich point clouds with other modalities, alongside methods that do not fit its main categories. Their usefulness depends on task and data conditions.

3.3 Adversarial deformation augmentation

  • Adversarial deformation perturbs point clouds to represent shape variation, but its use for augmentation remains limited and warrants further exploration.3D-VField uses sensor-ray vector fields for smooth instance deformation; its extension to segmentation redesigns the adversarial loss and can perturb intensity signals.Source: §3.3 · p. 20

3.4 Up-sampling augmentation

  • Up-sampling can augment sparse data with denser point clouds while retaining local semantics, but methods trained on synthetic data have performed poorly on complex real-world scans.Studies report improved classification on up-sampled ModelNet tests; one reconstruction study found networks extracted features more effectively from up-sampled point clouds.Source: §3.4 · p. 22

3.5 Completion augmentation

  • Completion augmentation creates incomplete training inputs and completes them, though voxelization may lose information and some methods depend on adjacent frames.The initial removal step may be unnecessary for already incomplete data; an end-to-end point encoder GAN is described as avoiding additional information loss.Source: §3.5 · pp. 22–23

3.6 Generation augmentation

  • Generation augmentation can produce varied synthetic instances while preserving global semantics, but scene-level generation described here cannot yet generate cars and pedestrians together.TreeGAN has high computational complexity; scene generation decodes codebook features for conditional or unconditional driving-scene generation.Source: §3.6 · p. 24

3.7 Multi-modal augmentation

  • Multi-modal augmentation uses complementary data such as images to densify point clouds, but acquiring and aligning other modalities can be difficult.The survey also notes that such augmentation may not outperform algorithms designed specifically for multi-modal analysis.Source: §3.7 · pp. 25–26

3.8 Others

  • Other specialized methods include scene mixing, local point swaps, student-network feature shuffling, and pseudo-labeled object insertion.Feature shuffling can split complete instances and produce localization bias; PseudoAugment removes low-confidence pseudo-labeled points before pasting objects into scenes.Source: §3.8 · pp. 26–27

4 Discussion and future work

The discussion describes how augmentation practices vary across point-cloud tasks and datasets, while noting that inconsistent evaluations limit fair comparisons. It identifies methodological gaps and future directions spanning more realistic, efficient, and standardized augmentation.

Current practice

  • Augmentation is commonly used in outdoor autonomous-driving detection and 3D-object classification, but operation choices differ by task: GT-sampling is seldom used for classification because it is unsuitable there.Indoor scene-understanding methods are not notably distinct from outdoor methods, with choices appearing to depend more on point-cloud characteristics than acquisition setting.Source: §4 · p. 27

Training uses

  • Augmentation supports both conventional training, where original and augmented data train networks, and consistency learning, where multiple augmented variants feed multiple networks.Consistency learning uses variants to encourage similar predictions under different input conditions.Source: §4 · p. 28, Figure 12 · p. 29

Evaluation

  • Reported studies find performance improvements after augmentation, but differences in datasets and baseline networks make fair comparisons across studies challenging.The survey examined reported before-and-after evaluations, commonly presented in ablation experiments; overall performance also depends on the downstream networks and benchmark datasets.Source: §4 · pp. 29–30

Future directions

  • Future work includes underexplored adversarial deformation, up-sampling, completion, and generation, as well as realistic augmentation ranges and temporal strategies for dynamic scenes.The survey also suggests exploring specialized augmentations in consistency learning and testing augmentations with state-of-the-art Transformer backbones.Source: §4 · pp. 31–32

Evaluation gaps

  • Few studies evaluate augmentation using consistent baseline networks and datasets across processing tasks, motivating shared evaluation approaches, metrics, or datasets.The survey also identifies a lack of universally accepted combinations of basic operations and recommends protocols suited to application domains, tasks, and datasets.Source: §4 · p. 31

Specialized methods

  • Specialized methods can supplement basic operations, but some are computationally expensive on large datasets or complex to reproduce, motivating efficient and plug-and-play implementations.The survey proposes exploring combinations such as GT-sampling with generation augmentation to extend generated instances to scene-level data.Source: §4 · pp. 31–32

5 Conclusion

The conclusion presents a taxonomy-based survey of point-cloud augmentation and summarizes patterns in the use of basic and specialized methods. It reports that augmentation improves downstream model performance while emphasizing that inconsistent evaluations constrain cross-study comparison.

5 Conclusion

  • The survey organizes point-cloud augmentation methods into a proposed taxonomy and discusses their strengths, limitations, and future research directions.The conclusion frames augmentation as a response to demand for larger and more diverse training datasets, especially when annotated point-cloud data are limited.Source: §5 · p. 32
  • Basic augmentation remains preferred across many perception tasks for its simplicity and adaptability, while GT-sampling with affine transformations is common in detection, occasional in segmentation, and rare in classification.Choices of basic-operation combinations often rely on trial experiments or practical experience, though algorithms for selecting operations and parameters have been explored.Source: §5 · pp. 32–33
  • Specialized methods target particular challenges: Mixup is widely used for 3D-object classification, while domain augmentation takes precedence in outdoor-autonomous-driving detection.The conclusion describes specialized methods as supplementary to basic methods and notes that their task-specificity and implementation complexity have not displaced basic methods broadly.Source: §5 · p. 33
  • The survey reports that augmentation improves deep-learning performance across downstream tasks, but inconsistent methods, networks, and datasets prevent quantitative cross-study comparisons.It calls for evaluation approaches, metrics, or datasets and a standard protocol for choosing augmentation methods according to dataset and task characteristics.Source: §5 · p. 33

Declaration of competing interest

Appendix A. Reported Performance in Downstream Tasks

Appendix A tabulates reported downstream-task performance using augmentation on commonly used benchmarks. The authors caution that the results do not isolate augmentation effects because studies use different networks and datasets.

Appendix scope

  • Appendix A summarizes detection, segmentation, and classification results on commonly used benchmarks, including KITTI, nuScenes, Waymo, SemanticKITTI, ShapeNet, ModelNet, and ScanObjectNN.The summary includes the augmentation methods used in the reported tasks.Source: Appendix A · p. 41

Interpretation

  • Appendix A results require caution because performance reflects augmentation methods as well as downstream networks and benchmark datasets.Studies typically test different augmentation and network combinations, selecting the combination with the highest task-specific score.Source: Appendix A · p. 41

Detection benchmarks

  • The appendix includes comparative tables for KITTI 3D detection test and validation benchmarks, KITTI BEV detection, and nuScenes and Waymo detection.The nuScenes table reports mAP and NDS, while the Waymo validation results use mAP and mAPH.Source: Table 1 · p. 41, Table 4 · p. 42, Table 2 · p. 42

L2 mAP\

The supplied appendix passages contain fragments of comparative results for detection and segmentation methods and identify a classification table covering ModelNet and ScanObjectNN. They do not support a unified performance ranking across these fragments.

L2 mAP\

  • The cited entries list MEGVII, PointAugmenting, and PointPillars with GT+T+R+S+F, Hu et al. with GT+T+R+F, and LidarAugment as Auto optimization.Source: mAPH · p. 42
  • The supplied entries also name specialized methods including Mix3D and PolarMix, alongside segmentation architectures such as KPConv and SPVCNN.Source: mAPH · p. 43
  • The classification comparison covers ModelNet and ScanObjectNN and uses overall accuracy as its evaluation metric.Its caption notes that unavailable results are marked with a dash and italic values come from replicated tests by others.Source: Table 6 · p. 43

Reference Used in Appendix A

Appendix A’s references cover point-cloud datasets and models for detection, segmentation, and classification, alongside methods for augmenting and generating point clouds.

Reference Used in Appendix A

  • The references include KITTI, nuScenes, Waymo Open Dataset, SemanticKITTI, ShapeNet, and a real-world point-cloud classification benchmark.They also cite multiple 3D object detection methods, including PointRCNN, SECOND, PointPillars, PV-RCNN, and SE-SSD.Source: Reference Used in Appendix A · p. 44
  • Augmentation references span local transformations, adversarial and rendering-based methods, automatic augmentation, and point-cloud mixing.The bibliography also includes point-cloud generation methods such as diffusion probabilistic models and Point Cloud GAN.Source: Reference Used in Appendix A · pp. 44–46
  • Segmentation references include methods for LiDAR semantic and panoptic segmentation, point-cloud learning, and self-supervised learning.Examples include PolarNet, Cylinder3D, Panoptic-PolarNet, PointNet, PointNet++, and CrossPoint.Source: Reference Used in Appendix A · p. 45

Related papers

Loading 2308.12113v5…