Y.A. Maniakov1, P.O. Arkhipov2, P.L. Stavtsev3

1–3 Orel Branch of Federal Research Center «Computer Science and Control» of the RAS (Orel, Russia)
1 maniakov_yuri@mail.ru; 2 arpaul@mail.ru; 3 pavelstavcev@gmail.com

Currently, the use of three-dimensional reconstruction is quite widespread in various fields, ranging from computer vision, architecture and geodesy to medicine.

A significant problem when using 3D reconstruction results is that they may contain defects, namely, be noisy and contain voids (holes). For passive methods, such as photogrammetry or Structure from Motion, the causes are related to object surface properties, lighting conditions, low image quality due to imperfect equipment, processing algorithms, etc. For active methods, for example, laser scanning, the appearance of noise and gaps is associated with atmospheric conditions, object surface properties, motion distortion (of the sensor or the scene), and registration/stitching errors.

One effective approach to address distortions in 3D reconstruction results is the use of deep learning technologies. This technology, in particular, requires the availability of a training dataset containing distortion data. In this regard, it is critically important to ensure the most realistic generation of artificial defects in the training datasets during the sample preparation stage. However, when using neural network methods, it is crucial to ensure maximally realistic generation of noise in 3D point clouds during the dataset preparation stage. This training strategy allows the network to better understand the real nature of recording errors, work robustly with different data sources, preserve accuracy and details on complex surfaces, and generalize well, performing effectively on test and real-world point clouds – something that cannot be achieved by training exclusively on synthetic «ideal» noise. Without this, networks learn on synthetic artifacts and generalize poorly to real input data.

Deep learning methods demonstrate the ability to fill gaps, suppress noise, and restore details. For instance, models such as PointCleanNet, PointProNets, and PointFilter are used for noise reduction, while Point Completion Network, Point Projection Network, and Rl-GAN-Net [8] are designed for point cloud completion (inpainting). However, all these methods do not handle the combination of heterogeneous defects – noise and voids simultaneously.

In most research concerning methods and algorithms for denoising 3D objects, for example, Noise2Score3D, NoiseTrans, «Deep Learning for 3D Point Cloud Enhancement», PointCleanNet, «Denoise and Contrast for Category Agnostic Shape Completion», the creation of defective models is limited to applying Gaussian noise to the original point cloud. This type of outlier generation is based on the normal distribution function, making it simple to implement and less predictable compared to methods of uniform noise distribution over the object.

The paper proposes a comprehensive model for generating defects for training sample objects, combining either the generation of noise or incompleteness, which allows the creation of realistic datasets for training a neural networks in order to eliminate the those defects at the stage of post-processing the results of three-dimensional reconstruction.

Maniakov Y.A., Arkhipov P.O., Stavtsev P.L. A defects generation model in point clouds for training datasets creation. Highly Available Systems. 2025. V. 21. № 3. P. 69−78. DOI: https://doi.org/10.18127/j20729472-202503-06 (in Russian)

1. Anette Eltner, Giulia Sofia. Structure from motion photogrammetric technique// Developments in Earth Surface Processes. 2020. V. 23. S. 1–24. DOI:10.1016/B978-0-444-64177-9.00001-1
2. Cai X., Jiang W., Xu R., Zhao W., Ma J., Liu S., Li Y. Analyzing Infrastructure LiDAR Placement with Realistic LiDAR. 2022. DOI:10.48550/arXiv.2211.15975
3. Bayrak O.C., Zhenyu M., Farella E.M., Remondino F., Uzarand M. ESTATE: a large dataset of under-represented urban objects for 3D point cloud classification. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. 2024. V. XLVIII-2-2024. P. 25–32.
4. Manyakov Yu. A., Arhipov P. O., Stavcev P. L. Issledovanie i analiz sushestvuyushih metodov postobrabotki oblakov tochek. Sistemy vysokoj dostupnosti. 2024. T. 20. № 3. S. 51−58. DOI: 10.18127/j20729472-202403-05
5. Marie-Julie Rakotosaona, Vittorio La Barbera, Paul Guerrero, Niloy J. Mitra, Maks Ovsjanikov. PointCleanNet: Learning to Denoise and Remove Outliers from Dense Point Clouds. arXiv:1901.01060v3. 2019. DOI: 10.48550/arXiv.1901.01060
6. Riccardo Roveri, A. Cengiz Öztirel, Ioana Pandele, Markus Gross. PointProNets: Consolidation of Point Clouds with Convolutional Neural Networks // EUROGRAPHICS. 2018. V. 37. № 2.
7. Dongbo Zhang, Xuequan Lu, Hong Qin, Ying He. Pointfilter: Point Cloud Filtering via Encoder-Decoder Modeling. arXiv:2002.05968v2. 2020. DOI: 10.48550/arXiv.2002.05968
8. Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, Martial Hebert. PCN: Point Completion Network. arXiv:1808.00671v3. 2019. DOI: 10.48550/arXiv.1808.00671
9. Weichao Wu, Zhong Xie, Yongyang Xu, Ziyin Zeng, Jie Wan. Point Projection Network: A Multi-View-Based PointCompletion Network with Encoder-Decoder Architecture // Remote Sens. 2021. V. 13. № 23. P. 4917. DOI: 10.3390/rs13234917
10. Muhammad Sarmad, Hyunjoo Jenny Lee, Young Min Kim. RL-GAN-Net: A Reinforcement Learning Agent Controlled GAN Network for Real-Time Point Cloud Shape Completion. arXiv:1904.12304v1. 2019. DOI: 10.48550/arXiv.1904.12304
11. Xiangbin Wei. Noise2score3d: Tweedie’s approach for unsupervised point cloud denoising. arXiv:2503.09283. 2025. DOI: 10.48550/arXiv.2503.09283
12. Hou G., Qin G., Sun M., Liang Y., Yan J. and Zhang Z. Noisetrans: Point cloud denoising with transformers, arXiv:2304.11812. 2023. DOI: 0.48550/arXiv.2304.11812
13. Siwen Quan, Junhao Yu, Ziming Nie, Muze Wang, Sijia Feng, Pei An, and Jiaqi Yang. Deep learning for 3d point cloud enhancement: survey. arXiv:2411.00857. 2024. 3 DOI: 10.48550/arXiv.2411.00857
14. Antonio Alliegro, Diego Valsesia, Giulia Fracastoro, Enrico Magli, Tatiana Tommas. Denoise and Contrast for Category Agnostic Shape Completion. arXiv:2103.16671v1. 2021. DOI: 10.48550/arXiv.2103.16671
15. Pedro Hermosilla, Tobias Ritschel, Timo Ropinski. Total Denoising: Unsupervised Learning of 3D Point Cloud Cleaning. arXiv:1904.07615v2. 2019. DOI: 10.48550/arXiv.1904.07615
16. Liu Y., Zou B., Xu J., Yang S., Li Y. Denoising for 3D point cloud based on regularization of a statistical low-dimensional manifold. Sensors (Basel Switzerland) 22. 2666 DOI: 10.3390/s22072666
17. Fang Z., Liu Y., Xu L., Shahed M.H., Shi L. Research on a 3D Point Cloud Map Learning Algorithm Based on Point Normal Constraints. Sensors 2024. 24, 6185. DOI: 10.3390/s24196185
18. Wang W., Liu X., Zhou H., Wei L., Deng Z., Murshed M., Lu X. Noise4denoise: Leveraging noise for unsupervised point cloud denoising. Computational Visual Media 2024. DOI: 10.1007/s41095-024-0423-3
19. Ziqian Wu, Jiahao Liu. Perlin noise and its improvements: A literature review. 2024. DOI:10.54254/2755-2721/77/20240437
20. Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, Xinyi Le. PF-Net: Point Fractal Network for 3D Point Cloud Completion. arXiv:2003.00410v1. 2020. DOI: 10.48550/arXiv.2003.00410
21. Chang A.X., Funkhouser T., Guibas L., Hanrahan P., Huang Q. and Li Z., Savarese S., Savva M., Song S., Su H., Xiao J., Yi L., Yu F. Shape­Net: An Information-Rich 3D Model Repository. arXiv:1512.03012v1. 2015. DOI: 10.48550/arXiv.1512.03012
22. Rockafellar R., Tyrrell R., Roger J-B Wets. Variational Analysis. Springer-Verlag. 2005. P. 117.
23. Wu T., Pan L., Zhang J., Wang T., Ziwei L., Dahua L. Density-aware Chamfer Distance as a Comprehensive Metric for Point Cloud Completion. arXiv. 2021. С. 4–16.