A Canadian Benchmark LiDAR Dataset for Urban Infrastructure and 3D Scene Understanding

Abstract

High-quality 3D perception is essential for autonomous vehicles, urban analytics, and the development of intelligent transportation systems. However, existing LiDAR datasets are limited in their representation of fine-grained roadway and pedestrian infrastructure, and geographic diversity, particularly for environments common in North American cities. This paper introduces YEG3D, a large-scale, point-wise annotated mobile laser scanning (MLS) dataset comprising more than 682 million points collected across 14 km of urban roadway in Edmonton, Canada. The dataset includes a fine-grained taxonomy of 18 semantic classes, with an emphasis on detailed pedestrian, cyclist, and roadway infrastructure rarely distinguished in existing benchmarks. We additionally present a comprehensive baseline evaluation using five state-of-the-art semantic segmentation models, including PointNet++, DGCNN, KPConv, KPConvX, and Point Transformer v3. Among the evaluated models, Point Transformer V3 achieves the strongest overall performance, attaining 81.8% overall accuracy, 46.2% mean Intersection over Union (mIoU), and 56.8% mean F1 score, outperforming all other architectures across both global and class-level metrics. Detailed confusion matrix analysis reveals that while large structural classes are segmented reliably, fine-grained elements such as markings, bike lanes, and crosswalks remain challenging due to sparsity, occlusion, and class imbalance. YEG3D provides a new foundation for advancing research in 3D semantic segmentation, urban perception, and infrastructure-aware autonomous systems, and will be expanded in future releases to broaden its geographic and semantic coverage.

Background & Summary

Autonomous vehicles (AVs) are poised to reshape the future of transportation, relying heavily on robust perception systems capable of understanding complex urban environments. Modern autonomous systems integrate advanced sensing modalities, such as LiDAR (Light Detection and Ranging), radar, and camera-based detectors, to support accurate localization, real-time obstacle detection, and continuous environmental awareness^1,2. Among these, LiDAR plays a particularly critical role by actively emitting near-infrared pulses and capturing dense three-dimensional (3D) point clouds that encode detailed geometric and physical properties of surrounding infrastructure. These 3D representations form the foundation for object recognition, behavior prediction, and adaptive decision-making in both mapped and unmapped settings. Beyond autonomous driving, high-quality 3D scene perception enables a wide range of applications, including high-definition (HD) mapping, smart-city development, and large-scale urban modelling^3–5. Several influential datasets, such as SemanticKITTI⁴, Toronto-3D³, nuScenes⁶, the Waymo Open Dataset⁷, and ONCE⁸, have advanced research in LiDAR-based semantic segmentation. However, many provide limited semantic granularity for roadway and pedestrian infrastructure, which constrains research requiring a detailed representation of the built environment.

To address this need, we introduce a new large-scale, point-wise annotated LiDAR dataset consisting of more than 682 million high-resolution mobile laser scanning (MLS) points collected across 14 km of roadway in Edmonton, Alberta, Canada. The dataset includes 18 semantic classes representing diverse urban infrastructure elements and is designed for continued expansion in future releases to further enhance geographic and semantic coverage. Figure 1 shows a sample scene from the dataset, illustrating 17 of the annotated classes with zoomed-in regions highlighting fine-grained structural details. In addition to presenting the dataset, we benchmark a suite of state-of-the-art 3D point cloud semantic segmentation models to establish a transparent and reproducible performance baseline. This work introduces a comprehensive benchmark alongside a large and continually growing dataset designed to support progress across a broad range of research areas, including vision-based perception, 3D data processing, geomatics, and remote sensing. By offering a unified testbed for assessing semantic segmentation and classification techniques on substantial MLS point clouds, the benchmark encourages methodological exchange and helps identify emerging challenges in understanding complex 3D urban environments.

Fig. 1.

Sample from the YEG3D dataset showing 17 of the 18 annotated classes, with zoom-In views highlighting detailed structures and class distinctions.

Related datasets

The development of AVs has driven significant advances in 3D point cloud collection and semantic segmentation, providing critical data for scene understanding, object recognition, and urban perception. Several large-scale LiDAR and photogrammetric datasets have been introduced to support training and benchmarking of semantic segmentation algorithms, offering diverse urban and indoor environments with annotated 3D point clouds. Table 1 provides a summary of prominent datasets, including sensor type, spatial coverage, number of points, and semantic classes.

Table 1.

Recent 3D point cloud datasets for semantic segmentation tasks.

Dataset	Description	Location	Date	Sensor	Length/Area	# Points	# Labelled Classes	# Evaluated Classes	LiDAR Data Format
Oakland9	Early mobile LiDAR dataset for urban scene understanding	Oakland, USA	2009	SICK LMS	~1.5 km	~1.6 M	44	5	x, y, z, label
iQmulus10	Dense LiDAR of an urban scene in France	Paris, France	2015	RIEGL	10 km	~300 M	22	8	x, y, z, label, reflectivity, num_echo, object ID
Semantic3D11	Benchmark for semantic segmentation of static scans	Central Europe	2017	Terrestrial laser scanner	N/A	4B	8	8	x, y, z, intensity, label, R, G, B
Paris-Lille-3D12	Urban mobile mapping dataset	Paris & Lille, France	2018	Velodyne HDL-32E	2 km	143.1 M	50	9	x, y, z, GPS time label reflectivity
SemanticKITTI4	Largest LiDAR sequences with point-wise labels	Karlsruhe, Germany	2019	Velodyne HDL-64E	39.2 km	> 4.5B	28	25	x, y, z, intensity, label
nuScenes6	Multi-sensor urban driving dataset	Boston and Singapore	2019	Velodyne HDL-32E + Radar + Camera	1000 scenes	~1.1B (LiDAR only)	23	23	x, y, z, intensity, label
Waymo Open Dataset7	Large-scale AV dataset with LiDAR	Multiple (USA)	2019	Multiple 360° LiDARs	1,150 segments	390 k frames	23	23	x, y, z, intensity, label
Toronto-3D3	Urban mapping in Toronto	Toronto, Canada	2020	Teledyne Optech Maverick	1 km	~78 M	8	8	x, y, z, R, G, B, intensity, GPS time, label
S3DIS23	Large-scale indoor scene dataset	Indoor areas in several buildings, USA	2016	Matterport	NA	273 M	13	13	x, y, z, label, R, G, B
ScanNet24	Large-scale RGB-D indoor dataset	Various indoor environments	2017	RGB-D	1513 scenes /205 M frames	242 M	20	20	x, y, z, label, R, G, B
DublinCity14	A city-scale, densely annotated airborne LiDAR dataset	Dublin, Ireland	2019	Aerial Laser Scanning	2 km2	260 M	13	13	x, y, z, label, R, G, B
SensatUrban16	A large-scale UAV photogrammetry point cloud dataset	United Kingdom	2020	UAV Photogrammetry	7.6 km2	~3.0 B	31	13	x, y, z, label, R, G, B
Campus3D19	UAV photogrammetric point cloud of NUS campus	National University of Singapore (NUS) Campus, Singapore	2020	DJI Phantom 4 Pro	1.58 km² (6 regions)	937.1 M	24	24	x, y, z, R, G, B, hierarchical label, instance label
ONCE8	Massive LiDAR dataset for AV research	China	2021	40-beam LiDAR	1 M frames, 15k Scenes	~1B	5	5	x, y, z, intensity, label
Hessigheim 3D15	A high-density UAV LiDAR and 3D mesh dataset for semantic analysis	Hessigheim, Germany	2021	Riegl VUX-1LR	N/A	74 M	11	11	x, y, z, label, R, G, B
Swiss3DCities17	UAV photogrammetric urban 3D point cloud dataset across three Swiss cities	Zurich, Zug & Davos, Switzerland	2021	High-resolution nadir & oblique camera array	2.7 Km2 (16 tiles)	226 M	5	5	x, y, z, R, G, B, Label
STPLS3D21	Large-scale aerial photogrammetric dataset combining real and procedurally generated synthetic point clouds for urban scene understanding	Los Angeles, CA, USA (real); procedurally generated synthetic environments	2022	DJI Phantom 4 Pro (real); Unreal Engine 4 + Bentley ContextCapture (synthetic)	17.27 km² (1.27 km² real + 16 km² synthetic)	N/A	18 (synthetic V3); 6 (real)	6 (evaluated on real subset)	x, y, z, R, G, B, Label
YTU3D18	Large-scale UAV photogrammetric urban dataset	Istanbul, Turkey	2023	DJI Matrice 300 RTK	2 Km2	1700 M	45	45	x, y, z, R, G, B, Label
HRHD-HK20	Photogrammetric benchmark dataset focused on high-rise, high-density urban scenes	Hong Kong, China	2023	UAV Photogrammetry	9.375 km² (150 tiles)	273 M	7	7	x, y, z, R, G, B, Label
WHU-Urban3D13	A large-scale annotated dataset of airborne and mobile LiDAR with panoramic imagery	China	2024	AS-900HL and HiScan-Z 3D sensors	14 km	200 M	19	19	x, y, z, intensity, label
TALD22	ALS benchmark dataset from Thiruvananthapuram urban environment	Thiruvananthapuram, Kerala, India	2025	Leica ALS70 LiDAR + Intergraph DMC II 230 Camera	9 km² (9 tiles)	123 M	4	4	x, y, z, R, G, B, return number, number of returns, scan angle rank, label
YEG3D (ours)	Large-scale urban outdoor lidar dataset in Edmonton	Edmonton, Canada	2025	HESAI XT32M1X sensor	~14 km	~682 M	18	18	x, y, z, intensity, label

Early urban LiDAR datasets, such as Oakland⁹, provided foundational data for mobile scene understanding, capturing ~1.5 km of urban streets with modest point density and limited annotated classes. Subsequently, iQmulus¹⁰ and Semantic3D¹¹ offered higher-resolution LiDAR scans and more extensive spatial coverage, ranging up to 10 km and billions of points, supporting research in mobile mapping, dense 3D reconstruction, and terrestrial scanning applications. Paris-Lille-3D¹² further contributed to urban mobile mapping with 2 km of dense scans and a richer set of 9 evaluated classes, enabling semantic analysis of roadway and street-level structures. More recent datasets such as WHU-Urban3D¹³ significantly extend these efforts by combining airborne and mobile LiDAR acquisitions with panoramic imagery, providing over 200 million annotated points across multiple cities and including additional attributes such as intensity and return information. Similarly, DublinCity¹⁴ offers a densely annotated airborne LiDAR dataset at city scale, with hierarchical labeling of urban elements from coarse to fine levels, supporting both segmentation and reconstruction tasks.

Recent large-scale datasets for autonomous driving, such as SemanticKITTI⁴, Toronto-3D³, nuScenes⁶, Waymo Open Dataset⁷, and ONCE⁸, provide extensive annotated LiDAR sequences collected from multiple cities and vehicles. For example, SemanticKITTI offers 39.2 km of LiDAR sequences with over 4.5 billion points and 25 semantic classes, forming a benchmark for sequence-based 3D perception. Toronto-3D focuses on Canadian urban streets, capturing ~78 million points over 1 km of road, while nuScenes and Waymo datasets integrate multi-modal sensors, including LiDAR, cameras, and radar, to enable multi-sensor fusion and comprehensive urban scene understanding. ONCE, collected with a 40-beam LiDAR, introduces one of the largest AV datasets in terms of point cloud volume, supporting large-scale research in autonomous perception.

Beyond vehicle-mounted LiDAR, high-resolution UAV-based and multi-modal datasets have also emerged. The Hessigheim 3D (H3D) dataset¹⁵ provides very high-density UAV LiDAR point clouds alongside textured 3D meshes, enabling fine-grained semantic segmentation and change detection analysis. In parallel, photogrammetry-based datasets such as SensatUrban¹⁶ and Swiss3DCities¹⁷ offer large-scale, uniformly dense urban point clouds reconstructed from aerial imagery, with SensatUrban reaching nearly three billion points across multiple cities. Additional datasets, including YTU3D¹⁸ and Campus3D¹⁹, further expand the diversity of photogrammetric benchmarks by providing large-scale urban and campus environments with rich semantic labels and, in some cases, hierarchical or multi-resolution annotations. HRHD-HK²⁰ specifically targets high-rise, high-density urban environments, addressing a gap in existing benchmarks that are predominantly based on low-rise European cities.

Recent efforts have also focused on increasing dataset diversity and scalability through synthetic and geographically varied data sources. STPLS3D²¹ introduces a hybrid dataset combining real and procedurally generated synthetic aerial photogrammetry point clouds, enabling large-scale training with automatically generated annotations. Meanwhile, datasets such as the TALD²² expand geographic representation by capturing complex tropical urban environments with high land-cover diversity, which are often underrepresented in existing benchmarks. Collectively, these datasets highlight a shift toward larger, more diverse, and more richly annotated urban scenes, spanning multiple sensing modalities including mobile, airborne, and photogrammetric approaches.

Indoor datasets, such as S3DIS²³ and ScanNet²⁴, provide complementary benchmarks for semantic segmentation within interior environments. These datasets consist of high-density scans, capturing millions of points in office buildings and residential interiors with 13–20 semantic classes. They have contributed to advancements in indoor scene reconstruction, object recognition, and semantic segmentation, complementing outdoor urban LiDAR datasets in the research scope. Overall, these datasets collectively span a wide range of environments, sensor configurations, and annotation schemes, forming a comprehensive foundation for the development and evaluation of 3D point cloud semantic segmentation methods. They have enabled progress in both algorithmic research and practical deployment of AV perception systems, serving as critical benchmarks for dense 3D scene understanding.

In addition to datasets developed for semantic segmentation, large-scale geospatial LiDAR collections such as the USGS 3D Elevation Program (3DEP)²⁵ provide extensive point cloud coverage across the United States. These datasets consist of high-density airborne LiDAR data primarily intended for elevation mapping, terrain modeling, and earth science applications. While 3DEP offers broad spatial coverage and high-quality geometric information, it does not include the fine-grained semantic annotations required for infrastructure-level scene understanding and benchmarking of segmentation models.

Point cloud semantic segmentation models

The advent of deep learning techniques has brought significant advancements to point cloud semantic segmentation. Numerous models leveraging these methods have been developed, achieving superior performance and setting new benchmarks in point cloud processing²⁶. In the literature, various models have been explored and tested to determine effective approaches for point cloud semantic segmentation. Among these, two early notable point-based models, PointNet and PointNet++, have gained significant attention for their robustness and effectiveness.

PointNet, introduced by Qi et al.²⁷, processes raw point clouds by independently extracting features for each point using multi-layer perceptrons (MLPs). It then aggregates these features using a symmetric function (max-pooling) to generate a global representation of the point cloud. This approach effectively addresses challenges like permutation invariance and rotation invariance. However, one limitation of PointNet is its inability to sufficiently capture local and neighborhood interactions due to its reliance on global feature extraction²⁸.

To address this limitation, Qi et al.²⁹ proposed PointNet++, an extension of PointNet that introduces a hierarchical network architecture. PointNet++ incorporates a farthest point sampling (FPS) algorithm to select points with maximum spatial separation as local region centers. These centers are then used to construct local regions through a grouping module. The backbone network of PointNet is applied recursively to learn features within these regions. This hierarchical approach enhances the model’s ability to capture local geometric features while maintaining the global structure. However, while PointNet++ improves the extraction of local features, its ability to fully capture directional and spatial relationships between points remains limited²⁸. Other point-based models are SO-Net³⁰, PointSIFT³¹, PointWeb³², ShellNet³³, RandLA-Net³⁴, PointASNL³⁵, and PointMLP³⁶.

Graph Convolutional Network (GCN)-based models are also widely used for point cloud semantic segmentation. One such model is the Dynamic Graph Convolutional Neural Network (DGCNN), introduced by Wang et al.³⁷. Unlike traditional graph convolutional networks, DGCNN features a dynamic graph that is updated after each network layer. This dynamic graph construction allows the model to perform edge convolution operations, extracting features from central points while capturing edge vectors between these points and their K nearest neighbors (KNN). While the overall architecture of DGCNN is similar to that of PointNet, it replaces multi-layer perceptrons with edge convolution layers. This modification enables DGCNN to identify local neighborhoods in Euclidean space and group similar features within the feature space, making it particularly effective for point cloud classification and segmentation tasks²⁶.

Convolution-based approaches adapt traditional convolution operations to irregular point cloud data. Kernel Point Convolution (KPConv)³⁸ replaces fixed grid kernels with learnable kernel points positioned in Euclidean space, enabling flexible and continuous convolution directly on point clouds. Extensions such as KPConvX³⁹ incorporate geometric kernel attention mechanisms that modulate convolution weights based on spatial context, improving the model’s ability to capture fine-grained geometric structures.

Additionally, voxel-based models have been employed for semantic segmentation of point clouds. Voxelization refers to the process of representing an object by dividing it into small, grid-like units called voxels, which are positioned closest to the object²⁸. For instance, VoxNet, developed by Maturana and Scherer⁴⁰, utilizes volumetric techniques to convert unstructured geometric data into a regular 3D grid, allowing standard CNN operations to be applied. A 3D CNN is then used to predict the class label directly from the occupancy grid. Although this method resolves the issue of point clouds’ lack of structure, it has several drawbacks, such as inefficiency in voxel grid organization due to point cloud sparsity, high memory usage during computations, extended training times, and potential loss of information. Researchers have since introduced various improvements to address these challenges and enhance the model’s performance²⁶. Other voxel-based models are SPLATNet⁴¹, LatticeNet⁴², SEGCloud⁴³, OctNet⁴⁴, O-CNN⁴⁵, and VV-Net⁴⁶.

Most recently, transformer-based models have become an effective tool for point cloud semantic segmentation. Initially developed for natural language processing (NLP) tasks such as sentiment analysis and machine translation, the Transformer architecture relies on self-attention mechanisms. These mechanisms enable the model to focus on different parts of the input data when making predictions. Given that point clouds consist of unordered, unstructured, and sparse points, the self-attention mechanism and feed-forward neural networks within Transformers are more suitable for point cloud processing than traditional CNNs. This makes transformer models highly efficient for dealing with the unique challenges of point cloud data²⁸.

An example of a transformer-based model is the Point Transformer, introduced by Zhao et al.⁴⁷. This model leverages self-attention layers specifically designed for point cloud data, constructing self-attention Point Transformer networks for effective processing. The network employs self-attention operators, integrating subtraction relations and adding trainable, parameterized position encoding to the attention vector and transformation features. Additionally, residual Point Transformer blocks enhance information exchange between local feature vectors, with the core Point Transformer facilitating the capture of complex spatial dependencies within the point cloud²⁸. Other transformer-based models are PAT⁴⁸, PCT⁴⁹, MLMST⁵⁰, FastPointTransformer⁵¹, and SAT⁵².

A more advanced transformer model, Point Transformer v3⁵³, focuses on scalability, introducing efficient neighbor-mapping strategies that significantly expand the receptive field while improving computational efficiency. The authors observe that model performance in 3D point cloud processing is influenced more by scale than by architectural complexity. PTv3 therefore simplifies neighbor search by replacing exact KNN with a serialized neighbor-mapping strategy tailored to structured point cloud layouts. This design dramatically expands the receptive field, from 16 to 1024 points, while improving computation speed (3 × faster) and memory efficiency (10 × more efficient) relative to PTv2. PTv3 achieves state-of-the-art results across more than 20 indoor and outdoor benchmarks, with further gains realized through multi-dataset joint training. Transformer-based architectures have also been adapted for point-wise classification and segmentation tasks. These models leverage self-attention and cross-attention mechanisms to capture relationships between distant points and heterogeneous feature spaces, offering strong scalability and robustness in complex and sparse point cloud environments^54,55.

Recent advances further improve the balance between local detail and global context. The Stratified Transformer⁵⁶ employs a stratified sampling strategy that densely samples nearby points and sparsely samples distant ones, enabling efficient modeling of long-range dependencies while maintaining computational efficiency. Similarly, PGFormer⁵⁷ introduces a group representation transformer (GRT) that combines group-based and feature-based attention with KPConv, enhancing local feature representation while preserving global context in large-scale urban scenes. In parallel, improvements to convolution-based methods have also been proposed. IPCONV⁵⁸ extends kernel point convolution (KPConv) by introducing multiple kernel shapes and a multi-shape neighborhood system to better capture geometric variability in airborne LiDAR data. Additionally, the LCA-KPConv model⁵⁹ integrates a local contextual attention mechanism into KPConv, adaptively reweighting kernel responses based on feature similarity. This approach improves segmentation performance, particularly for fine-grained structures and boundary delineation, while maintaining computational efficiency.

Research contributions

Despite the significant contributions of the currently available datasets, several limitations persist. Most existing resources emphasize coarse object-level semantics, such as vehicles, pedestrians, vegetation, and buildings, while offering limited detail on the fine-grained infrastructure elements required for infrastructure-aware perception and urban analytics, such as traffic signs, signals, and poles. For example, SemanticKITTI provides 25 classes but groups many roadway features into broad categories such as road, sidewalk, other-ground, and terrain, without distinguishing critical elements like crosswalks, bike lanes, or medians. Toronto-3D, while captured in a Canadian city, includes only eight evaluated classes and lacks detailed annotations for pedestrian infrastructure, traffic control devices, and road markings. Multi-modal AV datasets such as nuScenes and Waymo offer rich annotations for dynamic agents (e.g., cars, cyclists, construction vehicles, pedestrians) and selected static elements like poles or traffic signs, yet they do not distinguish between nuanced forms of ground-level infrastructure such as bike lanes or crosswalks. nuScenes additionally provides top-down rasterized semantic maps for drivable areas and sidewalks, but these maps are limited to an effective area of roughly 5 km² and were generated independently of the LiDAR point cloud⁷. Waymo provides a larger set of classes, including road, sidewalk, curb, and lane marker. However, the official Waymo split allocates only 150 test scenes out of 1150 total, a ~13% test proportion, which can limit benchmarking reliability, particularly for rare classes and for evaluating generalization across diverse infrastructure types.

Beyond semantic limitations, geographic and environmental diversity also remains constrained. Many datasets are collected in specific regions, such as Germany, Singapore, Boston, Paris, or Chinese metropolitan areas, and few capture northern urban environments with wide sidewalks, multi-modal corridors, or the types of pedestrian and active-transportation infrastructure common to North American cities. Notably, Toronto-3D is the only publicly available Canadian outdoor LiDAR dataset, but it covers only ~1 km of roadway with limited semantic granularity. Such gaps restrict the development of methods targeting urban design workflows that require fine-grained annotation of walkways, bike facilities, multi-use paths, curbs, medians, traffic control devices, and road markings.

In response to these shortcomings, our work contributes a high-resolution mobile LiDAR dataset with fine-grained, infrastructure-focused semantics across 18 classes, including Road, Other Road, Median, Sidewalk, Shared Pathway, Walkway, Bike Lane, Crosswalk, Marking, Pole, Sign, Signal, Fence, High Vegetation, Low Vegetation, Building, Vehicle, and Not Classified. Unlike prior datasets, this taxonomy explicitly separates pedestrian, cyclist, and multi-use infrastructure, enabling detailed analysis of active-transportation corridors and roadway geometry. Additionally, the dataset evaluation protocol uses a k-fold cross-validation framework, instead of the traditional fixed train–test split, to provide a more robust assessment of model generalization across diverse urban scenes. This strategy reduces potential spatial bias associated with single-split evaluation and enables more reliable assessment of fine-grained and spatially sparse infrastructure classes that may appear unevenly across different regions of the dataset. The dataset captures more than 14 km of urban roadway in Canada, expanding geographic diversity in large-scale MLS datasets. Together, these contributions address persistent gaps in semantic coverage, infrastructure detail, data split reliability, and environmental representation across existing point cloud benchmarks.

Methods

Data collection

The YEG3D dataset was acquired using an MLS system deployed across a large portion of Edmonton’s urban road network. In total, more than 120 km of roadway were scanned, of which 14 km has been fully annotated for this release. Additional segments will be annotated and incorporated in future versions of the dataset to further expand its geographic and semantic coverage. All data were collected directly by the authors in public roadway environments, and no human participants were involved in the data collection process. Mobile LiDAR data collection relies on the integration of multiple high-precision sensors, including a laser scanner, a Global Navigation Satellite System (GNSS) receiver, and an inertial measurement unit (IMU)⁶⁰. Together, these components provide reliable estimates of both sensor pose and 3D point locations, ensuring accurate reconstruction of the surrounding environment even under challenging urban conditions such as occlusions or satellite signal degradation⁶¹. The overall sensor configuration used in this study is illustrated in Fig. 2. Data in this study were captured using the HESAI XT32M1X, a high-performance mechanical LiDAR sensor designed for mobile mapping and autonomous navigation applications. The XT32M1X features a 32-channel rotating laser configuration, delivering a measurement range of 0.5–300 m with a ranging accuracy of ±1 cm. It provides a 360° horizontal field of view and a 40.3° vertical field of view, operating at 10 Hz⁶². Under these specifications, the sensor records up to 1.92 million points per second, producing dense and geometrically consistent 3D point clouds suitable for fine-grained infrastructure analysis, roadway geometry extraction, and semantic segmentation. To achieve accurate georeferencing, the MLS platform incorporated the PointPerfect simpleRTK2B Pro GNSS receiver, which provides high-precision positioning at 10 Hz⁶³. This enables robust trajectory estimation during vehicle motion, particularly when paired with real-time kinematic (RTK) correction services. Complementing the GNSS, the system also employed a MicroStrain 3DM-CV7-AHRS IMU operating at 500 Hz, supplying high-frequency inertial measurements essential for maintaining accurate pose estimation during GNSS outages or multipath interference.

Fig. 2.

Top-down representation of the mobile mapping platform, showing the configuration and relative positions of all integrated sensors.

Data processing and annotations

The raw output from the data collection process is a.pcap file, which serves as the foundational dataset for all subsequent stages of analysis. This file contains the unprocessed LiDAR point cloud data captured during the urban scanning process. To convert this raw data into a meaningful spatial representation, it was first processed through Simultaneous Localization and Mapping (SLAM) using the LidarView software. SLAM is a robust technique that constructs accurate 3D maps of the environment by continuously estimating the sensor’s position and orientation in space, while simultaneously registering the incoming scan data. This alignment and accumulation process is essential for producing a coherent and comprehensive point cloud that accurately reflects the scanned urban environment. By leveraging SLAM, individual LiDAR frames are synchronized and stitched together, enabling the creation of an aggregated point cloud that encompasses detailed spatial information across the entire surveyed area. Figure 3 illustrates a single LiDAR frame captured during data collection, highlighting the sparse and momentary nature of raw scans. Also, Fig. 3 shows the aggregated LiDAR file after the SLAM process, where multiple frames are integrated to form a detailed and continuous representation of the environment. Since the SLAM output is initially generated in a local coordinate system, the point cloud was subsequently transformed into global coordinates using the georeferencing methodology previously developed by the authors in⁶⁴, enabling accurate spatial alignment and measurement.

Fig. 3.

(a) A single LiDAR frame captured during data collection; (b) the aggregated LiDAR map produced after SLAM integration.

For post-processing, CloudCompare was used to analyze, clean, and refine the point cloud data. As an open-source platform widely recognized for its advanced visualization, filtering, and 3D editing capabilities, CloudCompare is well-suited for handling large-scale LiDAR datasets⁶⁵. The software was particularly effective for noise reduction, outlier removal, and scene segmentation, ensuring that the final point clouds used for annotation and analysis were both accurate and reliable. Manual semantic labeling was also performed within CloudCompare to generate the ground-truth annotations for the dataset. To manage memory usage during the annotation and training pipeline, the dataset was divided into 50-m segments, each of which required approximately two hours of manual labeling, depending on scene complexity. In total, the dataset resulted in 295 segmented files, amounting to almost 600 hours of manual annotation effort, excluding time spent on data collection and preprocessing. The workflow of data processing and annotations is illustrated in Fig. 4.

Fig. 4.

Data processing and annotation workflow.

Data Record

The YEG3D dataset was collected in Edmonton, Canada, covering approximately 14 km of urban neighborhoods and roadway corridors. The annotated portion comprises roughly 682 million 3D points distributed across 18 semantic classes, with data acquisition conducted in 2025. The overall survey spans more than 120 km of raw mobile LiDAR scans, from which the current release represents the fully annotated subset. Future versions of the dataset will incorporate additional annotated segments to further expand both geographic and semantic coverage. The dataset has been deposited in Zenodo⁶⁶ and assigned the https://doi.org/10.5281/zenodo.20561546.

To structure the dataset, the annotated region is partitioned into seven zones, as shown in Fig. 5, with each zone representing distinct geometric, functional, and environmental characteristics of the urban environment. Each zone includes a consistent set of files, as illustrated in Fig. 6, enabling users to reconstruct the vehicle trajectory, sensor observations, and semantic information. Specifically:

• Annotation files (.txt and.ply): Provided in both class-wise and scene-wise formats. The scene-level representations contain all semantic classes within a single labeled file, while the class-wise representations separate individual infrastructure categories into independent files. This multi-format organization improves compatibility with common visualization software, semantic segmentation pipelines, and large-scale 3D deep learning frameworks while reducing preprocessing requirements for downstream applications. The files contain global coordinates $(x,y,z)$, normalized LiDAR intensity values [0, 255], and semantic class labels.

• LiDAR data (.pcap): Raw packet captures recorded directly from the XT32M1X sensor.

• Trajectory files (.las): Georeferenced vehicle trajectory data used for spatial alignment and point cloud reconstruction.

• GNSS data (.csv): Include timestamp, latitude, longitude, altitude, and azimuth measurements collected during acquisition.

• IMU data (.csv): Include timestamp, roll, pitch, yaw, linear acceleration, and angular velocity measurements for motion estimation and trajectory stabilization.

Fig. 5.

Spatial layout of the seven annotated zones in the YEG3D dataset.

Fig. 6.

YEG3D data structure.

A key component of YEG3D is its fine-grained semantic taxonomy, designed to capture detailed roadways and pedestrian infrastructure often overlooked in existing datasets, as shown in Fig. 7. A summary of the number of points and distribution of labels in each section is shown in Table 2 and Fig. 8. The 18 object classes are defined as follows:

• Road: Paved roadway surfaces, including travel lanes, bus lanes, and on-street parking areas.

• Other Road: Driveways, off-street parking lots, and other paved vehicular areas not part of the main roadway.

• Median: Raised medians separating opposing or adjacent traffic flows.

• Sidewalk: Designated pedestrian sidewalks located adjacent to roadways.

• Shared Pathway: Multi-use paths intended for shared pedestrian and cyclist use, typically separated from the road.

• Walkway: Walkable spaces not classified as sidewalks, such as building entrances, plazas, and connector paths.

• Bike Lane: Marked or separated lanes designated for cyclists.

• Crosswalk: Marked pedestrian crossing areas across road surfaces.

• Pole: Vertical pole-like objects such as streetlights, utility poles, bollards, and signposts without attached signage.

• Sign: Traffic signs, including STOP, YIELD, regulatory, warning, and informational signs.

• Signal: Traffic signal heads and their supporting poles; when a pole contains both a signal and a sign, the sign portion is classified as Sign.

• Building: Exterior facades and structural surfaces of buildings.

• Fence: Fences, guardrails, and other narrow vertical barriers.

• High Vegetation: Trees, tall shrubs, and dense vertical vegetation.

• Low Vegetation: Grass, small plants, and other low-height vegetation.

• Marking: Road markings including lane lines, arrows, symbols, and other painted elements.

• Vehicle: Motor vehicles such as cars, vans, trucks, and buses.

• Not Classified: Points outside the defined taxonomy or areas that were intentionally left unlabeled.

Fig. 7.

Visualization of fine-grained roadway, pedestrian, and cyclist infrastructure features in the YEG3D dataset.

Table 2.

Number of points and semantic label distribution across the seven annotated zones of the YEG3D dataset (in thousands).

Zone	Bike Lane	Building	Crosswalk	Fence	High Vegetation	Low Vegetation	Marking	Median	Not Classified	Other Road	Pole	Road	Shared Pathway	Sidewalk	Sign	Signal	Vehicle	Walkway	Total
1	81	5,083	876	622	6,959	3,190	136	108	359	127	94	10,130	—	3,465	121	236	23	264	31,874
2	—	12,545	658	707	36,447	20,391	159	—	695	792	342	13,828	—	4,738	241	—	3,220	1,037	95,800
3	1,082	4,646	306	689	7,892	4,434	169	473	164	243	170	3,226	—	2,218	112	95	272	389	26,580
4	—	24,809	604	2,046	82,011	44,397	179	—	1,127	1,003	918	29,903	—	9,991	506	81	7,735	2,167	207,477
5	—	8,115	536	618	39,817	23,721	103	123	354	480	436	18,902	3,703	5,389	329	39	3,372	748	106,785
6	1,781	18,206	517	1,953	51,184	29,973	344	605	1,562	1,797	631	21,766	3,306	10,368	658	235	2,251	1,190	148,327
7	—	10,668	353	1,086	23,145	10,433	161	65	822	1,545	583	9,064	—	4,845	318	17	1,750	1,078	65,933

L1 = 770 m, L2 = 2,045 m, L3 = 510 m, L4 = 3,655 m, L5 = 2,511 m, L6 = 3,044 m, L7 = 1,444 m; where $L$ denotes the scan length for each zone.

Fig. 8.

Distribution of the 18 semantic classes in the YEG3D dataset.

Technical Validation

Semantic segmentation models

Semantic segmentation of point clouds aims to assign a semantic label to every point. With recent advances in 3D deep learning, this task can now be addressed using end-to-end neural networks³. In this study, we benchmark the dataset using five state-of-the-art semantic segmentation models. To provide a more robust evaluation of model generalization, the dataset was partitioned into seven spatial zones and evaluated using a 7-fold cross-validation framework, where one zone was used for testing in each fold while the remaining six zones were used for training. This strategy reduces spatial bias associated with fixed train–test splits and enables a more reliable assessment of performance across diverse urban environments and fine-grained infrastructure classes. For each model, class-level evaluation metrics were computed independently for every fold, and the final reported performance values were obtained by averaging the results across all seven folds. The reported results are intended primarily as baseline benchmarks, and further performance improvements could likely be achieved through additional architecture refinement and hyperparameter tuning. All models were trained and evaluated on an NVIDIA RTX 4000 Ada GPU with 20 GB of RAM, with batch sizes and other parameters in the models adjusted to fit the available memory. The following paragraph provides a brief overview of each model and its core contributions.

PointNet++²⁹: A hierarchical point-based model that captures local geometric features through iterative sampling and grouping. DGCNN³⁷: A graph-based model that dynamically constructs neighborhoods and applies EdgeConv operations to capture local and global relationships. KPConv³⁸: A convolution-based approach that applies learnable kernel points directly in Euclidean space for flexible feature extraction. KPConvX³⁹: An extension of KPConv incorporating geometric kernel attention to improve sensitivity to fine spatial structures. Point Transformer v3⁵³: A scalable transformer-based model that expands receptive fields efficiently through optimized neighborhood mapping.

Evaluation metrics

The semantic segmentation models were evaluated using Intersection over Union (IoU), Precision, Recall, F1 score for each class, as well as their mean values and the Overall Accuracy (OA). The metrics are defined as follows:

• Intersection over Union (IoU) for class i:

  $${\mathit{IoU}}_i=\frac{{\mathit{TP}}_i}{{\mathit{TP}}_i+{\mathit{FP}}_i+{\mathit{FN}}_i}$$
• Mean IoU (mIoU) for all classes:

  $$\mathit{mIoU}=\frac{1}{N}\sum _{i=1}^N{\mathit{IoU}}_i$$
• Recall for class i:

  $$\mathit{Recall}=\frac{{\mathit{TP}}_i}{{\mathit{TP}}_i+{\mathit{FN}}_i}$$
• F1 score for class i:

  $${F1\mathit{score}}_i=\frac{2\times \mathit{Precision}\times \mathit{Recall}}{\mathit{Precision}+\mathit{Recall}}$$
• Mean F1 score for all classes:

  $$\mathit{mean}F1\mathit{score}=\frac{1}{N}\sum _{i=1}^N{F1\mathit{score}}_i$$

• Overall accuracy (OA):

$$\mathit{OA}=\frac{\sum \mathit{TP}}{\mathit{Total\; Number\; of\; Points}}$$

Where:

• TP_i = the number of true positive predictions for the class i.

• FP_i = the number of false positive predictions for the class i.

• FN_i = the number of false negative predictions for the class i.

• N = the number of classes.

OA and mIoU provide an overall indication of the semantic segmentation quality, while the IoU for each class highlights the model’s performance on individual categories. Additionally, Recall, and F1 score offer further details: Recall measures the model’s ability to correctly identify all instances of a class, and the F1 score combines both measures to give a balanced assessment of class-level performance.

The technical validation across all evaluation metrics reveals clear performance differences among the evaluated semantic segmentation models on the YEG3D dataset as shown in Tables 3, 4. Point Transformer V3 delivers the strongest overall performance, achieving the highest overall accuracy (81.8%), mean IoU (46.2%), and mean F1 score (56.8%). It consistently outperforms the next-best model, KPConvX, which attains a mean IoU of 40.4% and a mean F1 score of 49.4%. KPConv ranks third with a mean IoU of 37.4% and mean F1 score of 46.2%, while DGCNN achieves a mean IoU of 32.0% and mean F1 score of 42.2%. PointNet++ exhibits the weakest performance across all metrics, achieving an overall accuracy of 74.1%, a mean IoU of 25.0%, and a mean F1 score of 32.4%, highlighting its limited ability to capture the complex geometric characteristics of large-scale urban mobile LiDAR scenes. Visual examples of model predictions and qualitative differences in segmentation outputs are presented in Fig. 9, illustrating how each architecture handles both dominant roadway features and fine-scale urban objects.

Table 3.

Per-class and mean IoU (%) of the evaluated semantic segmentation models on the YEG3D dataset.

Model	OA	Mean	Bike Lane	Building	Crosswalk	Fence	High Vegetation	Low Vegetation	Marking	Median	Not Classified	Other Road	Pole	Road	Shared Pathway	Sidewalk	Sign	Signal	Vehicle	Walkway
PointNet++29	0.741	0.250	0.000	0.547	0.000	0.068	0.751	0.652	0.005	0.000	0.010	0.055	0.165	0.678	0.011	0.315	0.551	0.012	0.215	0.000
DGCNN37	0.719	0.32	0.043	0.807	0.120	0.000	0.759	0.245	0.082	0.000	0.001	0.029	0.297	0.538	0.078	0.288	0.568	0.000	0.471	0.045
KPConv38	0.782	0.374	0.055	0.882	0.000	0.424	0.928	0.542	0.000	0.287	0.172	0.055	0.432	0.624	0.001	0.210	0.584	0.100	0.740	0.023
KPConvX39	0.805	0.404	0.000	0.881	0.031	0.381	0.931	0.672	0.005	0.063	0.138	0.085	0.602	0.662	0.024	0.318	0.718	0.240	0.715	0.080
Point Transformer V353	0.818	0.462	0.101	0.779	0.041	0.450	0.898	0.731	0.116	0.360	0.134	0.113	0.623	0.706	0.038	0.474	0.727	0.310	0.748	0.166

Table 4.

Per-class and mean F1 score (%) of the evaluated semantic segmentation models on the YEG3D dataset.

Model	OA	Mean	Bike Lane	Building	Crosswalk	Fence	High Vegetation	Low Vegetation	Marking	Median	Not Classified	Other Road	Pole	Road	Shared Pathway	Sidewalk	Sign	Signal	Vehicle	Walkway
PointNet++29	0.741	0.324	0.000	0.706	0.000	0.123	0.856	0.785	0.010	0.000	0.019	0.104	0.277	0.806	0.021	0.467	0.707	0.024	0.334	0.000
DGCNN37	0.719	0.422	0.074	0.893	0.210	0.000	0.863	0.391	0.150	0.000	0.003	0.056	0.454	0.693	0.121	0.443	0.721	0.000	0.612	0.085
KPConv38	0.782	0.462	0.090	0.937	0.000	0.589	0.963	0.687	0.000	0.401	0.288	0.102	0.595	0.756	0.001	0.344	0.734	0.171	0.795	0.043
KPConvX39	0.805	0.494	0.000	0.936	0.050	0.547	0.964	0.802	0.010	0.112	0.240	0.153	0.744	0.796	0.044	0.478	0.833	0.363	0.793	0.144
Point Transformer V353	0.818	0.568	0.149	0.872	0.073	0.612	0.946	0.842	0.204	0.456	0.232	0.196	0.763	0.823	0.060	0.639	0.839	0.435	0.813	0.280

Fig. 9.

Visual comparison of semantic segmentation results for the original point cloud, ground truth, and predictions from the semantic segmentation models.

Class-level trends indicate that large and frequently occurring classes consistently achieve the highest segmentation accuracy. Point Transformer V3 attains IoU scores of 89.8% for High Vegetation, 73.1% for Low Vegetation, and 70.6% for Road, while KPConvX achieves comparable performance with IoU values of 93.1%, 67.2%, and 66.2% for the same classes, respectively. Building and Vehicle classes also exhibit strong performance across the top-performing models, with KPConv and KPConvX exceeding 88% IoU for Buildings and Point Transformer V3 reaching 74.8% IoU for Vehicles. These results demonstrate that large, well-represented classes with distinctive geometric characteristics are learned effectively by all architectures.

In contrast, minority classes remain challenging across all models. Bike Lane, Crosswalk, Marking, Shared Pathway, Signal, and Walkway consistently record low IoU values due to their limited representation and complex spatial characteristics. Although Point Transformer V3 provides the strongest performance on most rare classes, IoU values remain relatively modest, including 10.1% for Bike Lane, 4.1% for Crosswalk, 11.6% for Markings, 3.8% for Shared Pathway, and 16.6% for Walkway. Several models completely fail to identify certain classes, producing 0% IoU scores for categories such as Crosswalks, Bike Lanes, Signals, or Markings. These findings underscore the persistent difficulty of segmenting small, sparse, and often occluded urban features in mobile LiDAR datasets.

Class imbalance remains a major factor influencing model performance. Dominant classes such as vegetation, roads, and buildings contain substantially more points than minority classes, allowing models to learn robust geometric representations. Thin vertical structures such as poles illustrate the importance of model capacity and contextual understanding. Point Transformer V3 and KPConvX achieve IoU scores of 62.3% and 60.2%, respectively, whereas PointNet++ reaches only 16.5%, indicating that advanced architectures with larger receptive fields and stronger feature aggregation mechanisms are more effective at capturing fine-scale objects. The square-root inverse-frequency weighting strategy used during training improved the representation of minority classes; however, the severe imbalance present in YEG3D continues to limit performance for categories with very few observations and substantial occlusion.

Overall, the results demonstrate that Transformer-based architectures, particularly Point Transformer V3, provide the most robust and consistent performance for large-scale outdoor curbside environments characterized by highly variable point densities and a mixture of large surfaces and fine-grained urban objects. KPConvX also achieves strong results and remains competitive across most classes. While DGCNN and KPConv perform well on common structural categories, they generally struggle to match the performance of Transformer-based approaches on rare, spatially irregular, and context-dependent classes that require stronger global feature modeling and long-range spatial reasoning.

Figure 10 presents the recall-based confusion matrix of the Point Transformer V3 model evaluated on Zone 6 of the YEG3D dataset, providing a detailed view of class-level prediction accuracy and misclassification patterns across all 18 semantic categories. This visualization highlights the model’s performance across all 18 semantic classes and reveals both its strengths and remaining challenges in urban point cloud segmentation. The model performs exceptionally well on structurally distinct and well-represented classes such as Building (94.9% recall), Vehicle (95.0%), High Vegetation (93.1%), Low Vegetation (90.7%), Road (89.5%), Sign (85.9%), and Median (85.9%). These classes benefit from abundant training samples, strong geometric characteristics, and clear spatial separation from neighboring categories, resulting in strong diagonal values and relatively limited confusion with other classes.

Fig. 10.

Recall-based confusion matrix for the Point Transformer V3 model evaluated on Zone 6 of the YEG3D dataset.

Moderate performance is observed for classes with greater geometric variability or contextual overlap. Pole achieves a recall of 77.2%, with most errors occurring through confusion with High Vegetation (7.3%), Building (4.4%), and Sign (3.2%). Fence attains 78.2% recall but is occasionally misclassified as Building (7.7%), and Low Vegetation (4.0%). Signal remains challenging, achieving only 33.5% recall, with substantial confusion toward Pole (18.9%), and Sign (14.0%). This pattern reflects the geometric similarity between traffic signals, poles, and mounted roadside infrastructure, particularly when partial occlusion limits the visibility of distinguishing features.

The most significant challenges occur among roadway and pedestrian infrastructure classes that exhibit similar geometric characteristics and occupy adjacent spatial locations. Sidewalk achieves only 43.7% recall and is frequently confused with Other Road (22.2%) and Low Vegetation (18.3%). Shared Pathway records a recall of 26.9%, with most instances predicted as Sidewalk (36.8%) or Other Road (28.3%). Walkway performs similarly, reaching 26.3% recall while being commonly classified as Low Vegetation (31.6%), Other Road (22.7%), or Sidewalk (17.5%). These results indicate that the model struggles to distinguish between different pedestrian-oriented surfaces when geometric differences are subtle and contextual cues are limited.

Several minority classes remain particularly difficult to segment. Crosswalk achieves only 3.9% recall and is overwhelmingly misclassified as Road (73.5%) or Other Road (19.2%). Bike Lane records a recall of 28.7%, with more than half of its points predicted as Road (53.6%). Marking is the most challenging class, achieving only 5.0% recall while 80.4% of marking points are classified as Road. These results demonstrate the difficulty of learning thin, sparse, and highly imbalanced surface features that share similar elevation and geometric properties with surrounding roadway surfaces.

Other Road remains another problematic category, achieving only 57.9% recall and exhibiting substantial confusion with Road (25.0%) and Low Vegetation (12.2%). Similarly, the Not Classified category achieves only 37.7% recall and is frequently confused with Low Vegetation (16.7%), Vehicle (14.3%), Fence (12.1%), and Building (10.0%), reflecting the heterogeneous nature of points assigned to this class.

Overall, the confusion patterns reveal that Point Transformer V3 effectively segments classes characterized by distinctive three-dimensional geometry, such as buildings, vegetation, vehicles, signs, and road surfaces. These results suggest that the model successfully leverages its self-attention mechanisms to capture both local geometric features and long-range spatial relationships within large-scale urban scenes. However, classes that depend on subtle contextual differences, functional semantics, or fine-scale surface markings remain difficult to distinguish. The highly asymmetric misclassification behaviour, where minority classes are frequently absorbed into dominant categories such as Road, Sidewalk, and Other Road, indicates that class imbalance continues to play a major role despite the model’s strong overall performance. For practical deployment in autonomous navigation, digital twins, and urban analytics, the model demonstrates high reliability for major structural elements but would benefit from additional strategies targeting rare roadway markings, cycling infrastructure, and pedestrian facilities.

Challenges in the YEG3D dataset

The YEG3D dataset presents several challenges arising from the complexity of outdoor curbside environments and the limitations of existing semantic segmentation frameworks. First, the intricate structure of urban streetscapes, characterized by heterogeneous objects, wide spatial extents, occlusions, and variable point densities, makes it difficult to extract consistent geometric features from the raw point cloud data. The sheer scale of the dataset further complicates manual annotation: large volumes of LiDAR points must be labeled with fine spatial precision, making the process time-consuming and labor-intensive even with specialized tools. Second, many established point cloud segmentation models (e.g., Pointnet++, DGCNN, KPConv) required substantial effort to deploy, as they rely on legacy libraries, CUDA versions, or compilation dependencies that are no longer supported. Adapting these models to modern frameworks involved reconfiguring environments, modifying data loaders, and resolving compatibility issues before meaningful benchmarking could be performed.

A further challenge stems from the highly imbalanced nature of the YEG3D dataset, a common issue in outdoor mobile-mapping scenarios. Several object categories, such as High Vegetation, Low Vegetation, Road, and Sidewalk, contain more points than smaller or more localized classes (e.g., Pole, Signal, Bike Lane, Fence, and Sign). This imbalance can bias model training toward majority classes, reduce recall for minority classes, and degrade overall segmentation performance. To mitigate this, class-balanced weighting was applied during training using a square-root inverse frequency formulation⁶⁷, which offers a smoother weighting scheme than full inverse frequency while still compensating for large disparities in class representation. For the experimental configuration in which Zone 6 was held out as the test set, the class weights were computed from the remaining training data and are presented in Table 5.

$$\begin{array}{ccc}{w}_i=\sqrt{\frac{N}{C\left(n_i+ϵ\right)}},& w_i^{\prime }=\frac{w_i}{\stackrel{\circledR }{w}},& \mathrm{where}\end{array}\stackrel{\circledR }{w}=\frac{1}{C}{\sum }_{j=1}^C{w}_j$$

where: N = total number of points, C = number of classes, $n_i$ = number of points belonging to class $i$, $ϵ$= small constant to avoid division by zero (1e-6), $w_i$ is the initial weight for class $i$, $w_i^{\prime }$ is the normalized weight, and $\stackrel{\circledR }{w}$ is the mean weight across all C classes. This normalization ensures that ${\sum }_{i=1}^C{w}_i^{\prime }=C,$ maintaining the relative importance between classes while keeping the overall loss magnitude comparable to the unweighted case.

Table 5.

Square-root inverse frequency class weights computed from the YEG3D training data using Zone 6 as the independent test set.

Class	Weight	Class	Weight
Bike Lane	1.713	Other Road	0.903
Building	0.228	Pole	1.159
Crosswalk	1.012	Road	0.20
Fence	0.769	Shared Pathway	0.960
High Vegetation	0.132	Sidewalk	0.334
Low Vegetation	0.179	Sign	1.448
Marking	1.940	Signal	2.701
Median	2.107	Vehicle	0.457
Not Classified	0.985	Walkway	0.775

Ablation Study and KPConvX Adaptation for YEG3D

KPConvX was originally developed and benchmarked on indoor datasets such as S3DIS and ScanNet, which contain dense, relatively uniform, and spatially constrained point distributions. In contrast, the YEG3D dataset consists of large-scale curbside MLS point clouds with substantially larger spatial extents, varying point densities, and irregular sampling patterns caused by vehicle motion, occlusions, and roadway geometry. These differences introduced challenges related to neighborhood construction, memory usage, and stable feature aggregation across complex outdoor scenes. To address these issues, several backbone and training parameters were adapted to better suit dense urban MLS data. Table 6 summarizes the primary configuration changes introduced for YEG3D.

Table 6.

Summary of the principal KPConvX configuration adaptations implemented for YEG3D.

Configuration Component	Baseline Setting	Updated Setting	Motivation/Impact
kp_radius	2.3	4	Increased receptive field size to improve contextual aggregation across large outdoor MLS scenes
radius_scaling	2.2	4	Improved multi-scale neighborhood coverage and feature propagation
subsample_size	0.02 m	0.04 m	Reduced computational cost and stabilized training on large-scale point clouds
GridSample (train)	0.02 m	0.04 m	Improved memory efficiency and reduced neighborhood instability
SphereCrop	40k pts	100k pts	Increased spatial context available during training and improved representation of roadway structures
Neighbor Limits	(12, 16, 20, 20, 20)	(25, 30, 35, 35, 35)	Improved neighborhood consistency and feature aggregation across scales
Scheduler	OneCycleLR (max LR 0.005)	OneCycleLR (max LR 0.001)	Increased spatial context available during training and improved representation of roadway structures
Point Limit	40k pts	100k pts	Improved optimization stability and reduced training divergence

One of the main modifications involved increasing the receptive field size to better capture contextual information across roadway environments. The kernel point radius (kp_radius) was increased from 2.3 to 4, while the radius scaling factor was increased from 2.2 to 4. These changes improved neighborhood coverage and enabled more stable feature aggregation across spatially heterogeneous urban scenes. Neighbor limits were also increased from (12, 16, 20, 20, 20) to (25, 30, 35, 35, 35) to support denser neighborhood construction and improve multi-scale feature propagation.

Several adjustments were further motivated by the computational demands of processing large MLS scenes. The subsampling size and training GridSample resolution were increased from 0.02 m to 0.04 m to reduce memory overhead and improve training stability. In parallel, the point crop limit was increased from 40k to 100k points, allowing the model to preserve larger roadway context and capture broader spatial relationships between roadway, pedestrian, and curbside infrastructure elements. The SphereCrop size was also increased from 40k to 100k points to improve spatial continuity during training and reduce fragmentation of urban structures. Additional modifications were introduced to improve optimization robustness and segmentation quality. The learning schedule was also adjusted by reducing the maximum learning rate from 0.005 to 0.001, which improved convergence stability during extended training on dense outdoor scenes.

The updated KPConvX configuration resulted in improved segmentation performance and greater training stability on the YEG3D dataset. Increasing the kernel point radius and radius scaling factor broadened the receptive field, which improved segmentation of spatially extended classes such as Road, and High Vegetation, where local neighborhoods alone are insufficient to capture class-discriminative structure. Enlarging the neighbor limits improved feature aggregation consistency across varying point densities, which is particularly relevant for classes with sparse or irregular sampling such as Pole, Sign, and Signal. Increasing the subsampling size and GridSample resolution reduced memory pressure and improved training stability, enabling longer and more consistent optimization across large MLS scenes. The corresponding increase in SphereCrop and point crop limits allowed the model to retain broader spatial context during training, which benefited classes whose identity depends on surrounding scene structure, such as Bike Lane, Crosswalk, and Sidewalk. Reducing the maximum learning rate improved convergence stability and reduced training divergence, which was observed as more consistent validation loss curves across epochs. Collectively, these adaptations were necessary to transfer KPConvX from its original indoor benchmarks to the large-scale, spatially heterogeneous conditions of the YEG3D dataset, and contributed to the overall segmentation performance reported in the paper.

Advantages and limitations

YEG3D offers several key advantages, including large-scale, high-resolution mobile LiDAR data with fine-grained, infrastructure-focused semantic annotations across 18 classes. The dataset captures detailed roadway, pedestrian, and curbside elements, such as crosswalks, bike lanes, and markings, that are often underrepresented or aggregated in existing benchmarks. Compared to airborne LiDAR, MLS provides significantly higher point density and improved capture of street-level features, enabling more precise representation of fine-scale urban infrastructure. This level of semantic granularity supports applications in transportation planning, roadway safety assessment, urban modeling, and intelligent mobility systems. In addition, its relatively balanced train–test split supports more robust and reliable benchmarking, particularly for evaluating generalization on fine-grained and spatially sparse classes. Also, the dataset reflects real-world sensing conditions, including occlusions, varying point densities, and sensor noise, providing a challenging and realistic testbed for evaluating model robustness and scalability.

Beyond semantic segmentation benchmarking, YEG3D has the potential to support broader research directions related to model generalization, representation learning, and cross-domain adaptation. Recent studies have shown that increasing the diversity of urban objects and sensing conditions can substantially improve the transferability and robustness of deep learning models across datasets and acquisition platforms. For example, Bayrak et al.⁶⁸ introduced the ESTATE dataset, which combines underrepresented urban objects collected from multiple datasets, sensor types, and geographic regions to improve 3D object classification performance. Their results demonstrated that increasing the diversity of sparse infrastructure objects, such as poles, traffic lights, and utility structures, can significantly improve deep learning robustness and cross-dataset generalization. Similarly, Bayrak et al.⁶⁹ later showed that integrating object-level datasets with semantic segmentation pipelines can improve segmentation performance for underrepresented urban classes. Given that YEG3D contains detailed curbside and roadway infrastructure captured under realistic MLS conditions, it may serve as a valuable complementary resource for improving representation diversity and supporting cross-domain urban perception research.

The dataset may also contribute to emerging research on self-supervised and foundation-model-based learning for 3D urban perception. Wu et al.⁷⁰ demonstrated that large and diverse point cloud collections can significantly improve representation quality in self-supervised learning frameworks, particularly under limited labeled data settings. Similarly, CitySeg⁷¹ highlighted the importance of multi-domain urban datasets for improving open-vocabulary segmentation and zero-shot generalization in city-scale point cloud understanding. Due to its dense street-level geometry, fine-grained semantic taxonomy, and realistic sensing artifacts, YEG3D may serve as a useful benchmark for evaluating representation learning, domain adaptation, and zero-shot transferability in MLS environments.

Another promising direction involves domain adaptation and parameter-efficient fine-tuning strategies. Carós et al.⁷² demonstrated that Low-Rank Adaptation (LoRA) can improve cross-domain segmentation performance while reducing computational cost and improving robustness to underrepresented classes. Since YEG3D exhibits substantial variability in point density, occlusions, roadway geometry, and urban structure, it provides a realistic environment for studying adaptation methods under challenging outdoor sensing conditions. Furthermore, its infrastructure-focused annotations may support future research on integrating semantic segmentation datasets with object-level classification benchmarks to improve the recognition of sparse and visually complex urban objects.

Despite its contributions, YEG3D has several inherent limitations stemming from its geographic specificity, class imbalance, and occlusion patterns caused by real-world urban conditions. The dataset is currently limited to a single Canadian city and therefore does not reflect the full diversity of urban forms, architectural styles, infrastructure configurations, and climatic conditions observed across different regions. While YEG3D provides detailed and fine-grained representation of street-level urban infrastructure, particularly roadways, pedestrian, and curbside elements, its scope is intentionally focused on these environments and does not extend to all types of large-scale or regionally varying infrastructure systems. Several classes remain underrepresented, reducing statistical robustness for rare-category evaluation and increasing the training difficulty for even the strongest models. Additionally, mobile LiDAR scanning introduces occlusions created by parked vehicles, vegetation, and building frontages, leading to incomplete observations that complicate both annotation and learning. These challenges, while reflective of real-world sensing conditions, limit the full generalizability of the dataset and highlight opportunities for broader data collection.

Future work will expand both the scope and richness of YEG3D along several directions. Additional annotated segments from the broader 120 km survey will be released incrementally, improving the representation of rare classes (e.g., crosswalks, bike lanes, and road markings) and increasing environmental diversity across neighborhoods, roadway types, and seasonal conditions. Expanding data collection to additional Canadian cities will further address geographic specificity and support model generalization across varied climatic and infrastructural contexts. The semantic taxonomy will be refined to include additional curbside elements, pedestrian amenities, and roadway furniture, enabling more detailed infrastructure analysis. We will also explore semi-automated annotation pipelines to reduce labeling effort and improve consistency. Multi-platform data collection, including scooter-mounted systems, will be investigated to mitigate occlusions inherent in vehicle-based LiDAR and to support multi-scan fusion and registration. Incorporating complementary modalities (e.g., RGB imagery, thermal data, and radar) and extending the dataset temporally will further enable research in multi-modal perception, change detection, and dynamic scene understanding. YEG3D also provides a strong foundation for urban digital twin applications. High-fidelity semantic point clouds can support urban planning, infrastructure monitoring, and smart city decision-making. Future releases will prioritize interoperability with open data standards and simulation platforms. Key applications include infrastructure condition assessment, accessibility analysis, and temporal change detection. Collaboration with municipal partners will further align dataset development with real-world urban analytics needs.

Author contributions

The authors confirm their contributions to the paper as follows: study conception and design: A.S., F.B. and K.B.; data collection: A.S., F.B.; analysis and interpretation of results: all authors; draft manuscript preparation: all authors. All authors reviewed the results and approved the final version of the manuscript.

Funding

We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [funding reference number ALLRP 586279 - 23].

Data availability

The dataset has been deposited in Zenodo⁶⁶ and assigned the https://doi.org/10.5281/zenodo.20561546.

Code availability

No custom code was used to generate the data described in the manuscript.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41597-026-07830-y.