LiDAR point-cloud segmentation is an important problem for many applications. For large-scale point cloud segmentation, the de facto method is to project a 3D point cloud to get a 2D LiDAR image and use convolutions to process it.

UA_L-DoTT (University of Alabama’s Large Dataset of Trains and Trucks) is a collection of camera images and 3D LiDAR point cloud scans from five different data sites. Four of the data sites targeted trains on railways and the last targeted trucks on a four-lane highway. Low light conditions were present at one of the data sites showcasing unique differences between individual sensor data. The final data site utilized a mobile platform which created a large variety of view points in images and point clouds. The dataset consists of 93,397 raw images, 11,415 corresponding labeled text files, 354,334 raw point clouds, 77,860 corresponding labeled point clouds, and 33 timestamp files. These timestamps correlate images to point cloud scans via POSIX time. The data was collected with a sensor suite consisting of five different LiDAR sensors and a camera. This provides various viewpoints and features of the same targets due to the variance in operational characteristics of the sensors. The inclusion of both raw and labeled data allows users to get started immediately with the labeled subset, or label additional raw data as needed. This large dataset is beneficial to any researcher interested in machine learning using cameras, LiDARs, or both.

The full dataset is too large (~1 Tb) to be uploaded to Mendeley Data. Please see the attached link for access to the full dataset.

NPM3D (https://npm3d.fr/paris-carla-3d) consists of mobile laser scanning (MLS) point clouds collected in four different regions in the French cities of Paris and Lille, where each point has been annotated with two labels: one that assigns it to one out of 10 semantic categories and another one that assigns it to an object instance. When inspecting the data, we found 9 cases where multiple tree instances had not been separated correctly (i.e., they had the same ground truth instance label). These cases were manually corrected using the CloudCompare software (https://www.cloudcompare.org), and 35 individual tree instances were obtained. Our variant of the dataset with 10 semantic categories and enhanced instance labels is publicly available.

datasets used in the experiment

LiDAR scans of 24 avocado trees from several years. Each point cloud has been annotated with two labels: one label for leaf (0) vs trunk (1) matter, and another for which tree each point belongs to: ground (0), center tree (1), north tree (2), south tree (3), uncategorised (4). Each point cloud is stored as a binary file with the following format (ui = 4-byte unsigned integer, d = 8-byte double): 3d,2ui,d Each line in the file has the fields: x,y,z,matter_label,tree_label,height XYZ is in North-East-Down orientation. The data was created in, and is easy to view using, ACFR comma/snark open-source tools (https://github.com/acfr/comma/wiki)

Also included in the dataset is a list of trunk points in simple CSV format with tree IDs ; there are other trees in that list as well, but all trees represented in this dataset are in there.

Error analysis of point cloud extraction method for buildings.

Toronto-3D is a large-scale urban outdoor point cloud dataset acquired by an MLS system in Toronto, Canada for semantic segmentation. This dataset covers approximately 1 km of road and consists of about 78.3 million points. Here is an overview of the dataset and the tiles. The approximate location of the dataset is at (43.726, -79.417).

The XY coordinates are stored in UTM format. The Y coordinate may exceed decimal digits in float type commonly used in point cloud processing algorithms. Directly read and process the coordinates could result in loss of detail and wrong geometric features.

I set a UTM_OFFSET = [627285, 4841948, 0] to subtract from the raw coordinates. You may use any other numbers to reduce number of digits.

Example of potential issues during grid_subsampling operation used in KPConv and RandLA-Net: both subsampled to grid size 6cm

The proposed dataset, termed PC-Urban (Urban Point Cloud), is captured with an Ouster LiDAR sensor with 64 channels. The sensor is installed on an SUV that drives through the downtown of Perth, Western Australia (WA), Australia. The dataset comprises over 4.3 billion points captured for 66K sensor frames. The labelled data is organized as registered and raw point cloud frames, where the former has a different number of registered consecutive frames. We provide 25 class labels in the dataset covering 23 million points and 5K instances. Labelling is performed with PC-Annotate and can easily be extended by the end-users employing the same tool.The data is organized into unlabelled and labelled 3D point clouds. The unlabelled data is provided in .PCAP file format, which is the direct output format of the used Ouster LiDAR sensor. Raw frames are extracted from the recorded .PCAP files in the form of Ply and Excel files using the Ouster Studio Software. Labelled 3D point cloud data consists of registered or raw point clouds. A labelled point cloud is a combination of Ply, Excel, Labels and Summary files. A point cloud in Ply file contains X, Y, Z values along with color information. An Excel file contains X, Y, Z values, Intensity, Reflectivity, Ring, Noise, and Range of each point. These attributes can be useful in semantic segmentation using deep learning algorithms. The Label and Label Summary files have been explained in the previous section. Our one GB raw data contains nearly 1,300 raw frames, whereas 66,425 frames are provided in the dataset, each comprising 65,536 points. Hence, 4.3 billion points captured with the Ouster LiDAR sensor are provided. Annotation of 25 general outdoor classes is provided, which include car, building, bridge, tree, road, letterbox, traffic signal, light-pole, rubbish bin, cycles, motorcycle, truck, bus, bushes, road sign board, advertising board, road divider, road lane, pedestrians, side-path, wall, bus stop, water, zebra-crossing, and background. With the released data, a total of 143 scenes are annotated which include both raw and registered frames.

The SemanticTHAB dataset is a large-scale dataset designed for semantic segmentation in autonomous driving. It contains 4,750 3D LiDAR point clouds collected from urban environments. The dataset includes labeled point clouds with 20 semantic classes, such as road, car, pedestrian, and building. It provides ground truth annotations for training and evaluating semantic segmentation algorithms, offering a real-world benchmark for 3D scene understanding in self-driving car applications. The dataset is desinged to extent the SemanticKITTI benchmark by scans of a modern high resolution LiDAR sensor (Ouster OS2-128, Rev7).

Supervised training of a deep neural network for semantic segmentation of point clouds requires a large amount of labelled data. Nowadays, it is easy to acquire a huge number of points with high density in large-scale areas using current LiDAR and photogrammetric techniques. However it is extremely time-consuming to manually label point clouds for model training. We propose an active and incremental learning strategy to iteratively query informative point cloud data for manual annotation and the model is continuously trained to adapt to the newly labelled samples in each iteration. We evaluate the data informativeness step by step and effectively and incrementally enrich the model knowledge. We use the airborne laser scanning point clouds captured over the Rotterdam central to evaluate our proposed method. Date Submitted: 2020-12-16

Producing reliable log volume data is an essential feature in an effective wood supply chain, and LiDAR sensing, supported by portable platforms, is a promising technology for volume measurements. Computer-based algorithms like Poisson interpolation and Random Sampling and Consensus (RANSAC) are commonly used to extract volume data from LiDAR point clouds, and comparative studies have tested these algorithms for accuracy. To extract volume data, point clouds require several post-processing steps, while their outcome may depend largely on human input and operator decision. Despite the increasingly number of studies on accuracy limits, no paper has addressed the reliability of these procedures. This raises at least two questions: (i) Would the same person, working with the same data and using the same procedures get the same results? And (ii) How much would the results deviate when different people process the same data using the same procedures? A set of 432 poplar logs placed on the ground and spaced about 1 m apart, was scanned by a professional mobile LiDAR scanner in groups; the first 418 logs were then individually scanned using an iPhone-compatible app, with the remainder being excluded from this part of the study due to field time constraints and all the logs were manually measured to get the reference biometric data. Three researchers with different experiences processed the datasets produced by scanning twice, following a protocol that included shape reconstruction and volume calculation using Poisson interpolation and RANSAC algorithm for cylinders and cones. The intra- and inter-rater reliability were evaluated using a comprehensive array of statistical metrics. The results show that the most reliable estimates correlate with a greater experience. The Cronbach’s alpha metric at the subject level was high, with values of 0.902–0.965 for the most experienced subject, and generally indicated moderate to excellent intra-rater reliabilities. Moreover, working with Poisson interpolation and RANSAC cylinder shape reconstruction, respectively, indicated a moderate to excellent reliability. For the Poisson interpolation algorithm, the Intraclass Correlation Coefficient (ICC) ranged from 0.770 to 0.980 for multi-log datasets, and from 0.924 to 0.972 for single log datasets. For the same type of input datasets, the ICC varied between 0.761 and 0.855 and from 0.839 to 0.908 for the RANSAC cylinder, and from 0.784 to 0.869 and 0.843 to 0.893 for the RANSAC cone shape reconstruction algorithms, respectively. These values indicate a moderate to excellent inter-rater reliability. Similar to Cronbach’s alpha, the Root Mean Square Error (RMSE) was related in magnitude to the ICC. The results of this study indicate that, for improved reliability and efficiency, it is essential to automate point cloud segmentation using advanced machine learning and computer vision algorithms. This approach would eliminate the subjectivity in segmentation decisions and significantly reduce the time required for the process.

The dataset is an annotated point cloud in ASPRS LAS v1.2 format, which is annotated with different classification numbers representing six different road markings, including lane markings (1), pedestrian crosswalk and text (2), bike (3), left arrow (4), right arrow (5), straight arrow (6), and others (0). The point cloud dataset was obtained using Oregon Department of Transportation current mobile lidar system (Leica Pegasus:Two). The data were georeferenced in the supporting software for the Leica Pegasus:Two by Oregon DOT. The authors processed the data to extract the road markings using the road marking extraction tool (Rome2) developed in this Pactrans research.

This LiDAR point cloud dataset is collected with a research platform of Finnish Geospatial Research Institute (FGI), called Autonomous Research Vehicle Observatory (ARVO). The dataset was collected with Velodyne VLS-128 Alpha Puck LiDAR, 7th of September 2020 in a suburban environment in the area of Käpylä in Helsinki, the capital of Finland. The environment in the dataset consists of a straight two-way asphalt street, called Pohjolankatu, which starts from a larger controlled intersection at the crossing of Tuusulanväylä (60.213326° N, 24.942908° E in WGS84) and passes by three smaller uncontrolled intersections until the crossing of Metsolantie (60.215537° N, 24.950065° E). It is a typical suburban street with tram lines, sidewalks, small buildings, traffic signs, light poles, and cars parked on both sides of the streets. To collect a reference trajectory and to synchronize the LiDAR measurements, we have used a Novatel PwrPak7-E1 GNSS Inertial Navigation System (INS).

The motion distortion of each individual scan has been corrected with a postprocessed GNSS INS trajectory and the scans have been registered with Normal Distributions Transform (NDT). Each point is provided with a semantic label probability vector and the final point cloud is averaged with a 1 cm voxel filter.

The steps to create this preprocessed dataset have been described in more detail in the article "Towards High-Definition Maps: a Framework Leveraging Semantic Segmentation to Improve NDT Map Compression and Descriptivity" published in IROS 2022. However, the number of points in each semantic segment in Table I in Section IV-A are different. The correct values are shown in the table below. This does not affect the results.

TABLE I: RandLA-Net classified dataset label proportions.


    Semantic label
    No. of points
    % of all
    % of used


    Ground
    14,206,060
    32.3
    50.3


    Building
    7,782,757
    17.7
    27.6


    Tree Trunk
    3,736,775
    8.5
    13.2


    Fence
    2,201,851
    5.0
    7.8


    Pole
    206,983
    0.5
    0.7


    Traffic Sign
    85,316
    0.2
    0.3


    Labels used here
    28,219,742
    64.1
    100.0


    Others
    15,821,962
    35.9



    Total
    44,041,704
    100.0

Monitoring and preserving forests is becoming increasingly important due to the escalating effects of climate change and threats of deforestation. In the domain of forest science, three-dimensional data acquired through remote sensing technology has gained prominence for its ability to provide deep insights into the complex nature of forest environments. The process of identifying and segmenting individual trees in three-dimensional point clouds is a crucial yet challenging prerequisite for many forest analyses such as the classification of tree health and species. Tree segmentation is currently dominated by classical approaches that often rely on the forest’s canopy height model to identify tree crowns, but with limited success in complex environments and in particular areas underneath the canopy. Recent deep learning models are adept at performing instance segmentation on point clouds, but the performance of these models relies on the quantity and quality of training data. The difficulty of obtaining forest data owing to the cost of technology and annotation process hinders the development of neural networks for tree segmentation in forest point clouds. In this thesis, a scalable workflow is presented to produce arbitrarily large quantities of synthetic forest point clouds, and its effectiveness in deep learning is demonstrated. It is shown that by applying large amounts of synthetic forest data to pretrain neural networks, the individual tree segmentation performance in synthetic and real forests is significantly improved, outperforming classical segmentation methods. It is concluded that this workflow is effective at producing large quantities of realistic forest data, and its incorporation in deep learning fosters progress in tackling tree segmentation in forest point clouds. Its efficiency and scalability further indicate its potential for the development of frameworks, benchmarking systems, high throughput data analysis, and other analytical tasks.

The spectral clustering method has notable advantages in segmentation. But the high computational complexity and time consuming limit its application in large-scale and dense airborne Light Detection and Ranging (LiDAR) point cloud data. We proposed the Nyström-based spectral clustering (NSC) algorithm to decrease the computational burden. This novel NSC method showed accurate and rapid in individual tree segmentation using point cloud data. The K-nearest neighbour-based sampling (KNNS) was proposed for the Nyström approximation of voxels to improve the efficiency. The NSC algorithm showed good performance for 32 plots in China and Europe. The overall matching rate and extraction rate of proposed algorithm reached 69% and 103%. For all trees located by Global Navigation Satellite System (GNSS) calibrated tape-measures, the tree height regression of the matching results showed an value of 0.88 and a relative root mean square error (RMSE) of 5.97%. For all trees located by GNSS calibrated total-station measures, the values were 0.89 and 4.49%. The method also showed good performance in a benchmark dataset with an improvement of 7% for the average matching rate. The results demonstrate that the proposed NSC algorithm provides an accurate individual tree segmentation and parameter estimation using airborne LiDAR point cloud data.

Welcome to DUT Sewer3D Semantic Segmentation Dataset !

Motivation and Background

Urban sewer pipelines, as the critical guarantors of urban resilience and sustainable development, undertake the task of sewage disposal and flood prevention. However, in many countries, the most municipal sewer systems have been in service for 60 to 100 years, with the worst condition rating (D+) evaluated by ASCE.

As laser scanning is fast becoming the state-of-the-art inspection technique for underground sewers, semantic segmentation of pipeline point clouds is an essential intermediate step for pipeline condition assessment and digital twinning. Currently, similar to other building structures, the scarcity of real-world point clouds has hindered the application of deep learning techniques for automated sewer pipeline semantic segmentation.

Goal

We provided a high-quality, realistic, semantically-rich public dataset named "**Sewer3D Semantic Segmentation**" (S3DSS), including 800 synthetic scans and 500 real-world scans, for point cloud semantic segmentation in sewer pipeline domain, for which there are no public datasets in the past. S3DSS contains over 917 million points with 8 categories of common sewer defects. We hope it can be a starting point for benchmarking developed approaches to promote deep learning research on point cloud of sewer pipeline defects.

Content and Annotations

The two sub-datasets were obtained in the following way.

The real point cloud data were captured in laboratory scenarios using a FARO Focus S laser scanner. We used two prototype reinforced concrete sewer pipes to create most of the defect scenes. However, for misalign and displace defects that are difficult to operate with concrete pipes, we used two steel pipes which were well-designed to simulate. A total of 500 real scans were collected.

The synthetic point cloud data were obtained by our automated synthetic data generator in Unity3D. The introduction to the synthetic point cloud data generation methodology can be found in our paper. We generated 800 scans of sewer defect scenes. If you need more data, please contact Minghao Li (liminghao@dlut.edu.cn). In S3DSS, 8 common defect classes are used which includes:

• Normal
• Spalling
• Blockage
• Corrosion
• Misalignment
• Deposit
• Displacement
• Rubber Ring

Acknowledgements

This work was supported by the National Key R & D Program of China (Grant No. 2022YFC3801000) and the National Natural Science Foundation of China (Grant No. 52479118). We also thank Haurum et al. for sharing their great work "Sewer Defect Classification using Synthetic Point Clouds" as a reference for this work.

Citation

• Please cite this paper if you find this dataset useful:

【M. Li, X. Feng, Z. Wu, J. Bai, F. Yang, Game engine-driven synthetic point cloud generation method for LiDAR-based defect detection in sewers, Tunnelling and Underground Space Technology 163 (2025) 106755. https://doi.org/10.1016/j.tust.2025.106755.】

【Z. Wu, M. Li, Y. Han, X. Feng, Semantic segmentation of 3D point cloud for sewer defect detection using an integrated global and local deep learning network, Measurement 253 (2025) 117434. https://doi.org/10.1016/j.measurement.2025.117434.】

• Previous Publications:

M. Li, X. Feng, Q. Hu, 3D laser point cloud-based geometric digital twin for condition assessment of large diameter pipelines. Tunnelling and Underground Space Technology 142 (2023) 105430.

A large-scale synthetic dataset for LiDAR semantic segmentation, consisting of 13 LiDAR point cloud sequences with 198,396 scans in total.

WildScenes is a large-scale 2D and 3D semantic segmentation dataset containing both labelled images and lidar point clouds, in natural environments. The data was collected from two natural environments in Brisbane, Australia across multiple revisits. Our release includes 2D images, 2D annotated images, 3D point cloud submaps, 3D annotated point cloud submaps, alongside accurate 6-DoF poses. Lineage: The data was collected using a handheld sensor payload consisting of a spinning lidar sensor mounted at an angle of 45 degrees to maximise the field of view, a motor, encoder, an IMU, and four cameras. For each collected sequence we use the Wildcat slam system to create an accurate 6DoF estimation of the pose of the sensor and to process the lidar data into a globally registered map, from which we produce our submaps. The images we collected were manually annotated with per-pixel annotations and label transfer, using Paintcloud, was used to project 2D annotations into our 3D lidar maps.

A Novel Urban Biological Parameter Estimation Method Based on LiDAR Point Cloud Single-tree Segmentation

Semantic-KITTI dataset for 3D LiDAR point cloud semantic segmentation